What’s New

Recent changes to the Encyclopedia.

2026-05-16

What’s New

• New article: Hard Coding – the antipattern of baking values into source that should live somewhere a reader, operator, or future agent can change.
• New article: Architecture Astronaut – the antipattern of designing at an altitude so high that the abstractions stop touching any real problem.
• Improved: Edited the Copy-Paste Programming antipattern for prose quality, giving it a distinct rhetorical shape from the sibling Cargo Cult Programming entry and tightening the Why It Happens and Way Out sections.
• Improved: Edited the Cargo Cult Programming antipattern for prose quality, adding a Feynman etymology paragraph for accessibility and giving its illustrative scenarios distinct shape from the sibling Architecture Astronaut entry.
• Sources: Added external citation links to the How This Book Writes Itself article – seven canonical URLs covering Alexander’s A Pattern Language and The Timeless Way of Building, the Gang of Four’s Design Patterns, Wiener’s Cybernetics, and Brian Cantwell Smith’s 1982 MIT dissertation on procedural reflection.

Metrics

• Total articles: 259
• Coverage: 259 of 263 proposed concepts written (98.5%)
• Articles changed since last checkpoint: 2 new, 2 edits, 1 sources

2026-05-12

What’s New

• New article: Cargo Cult Programming – the trap of copying patterns, frameworks, and generated code shapes without understanding the reason they worked in their original context.
• New article: Copy-Paste Programming – the trap of duplicating code or rules across many files instead of giving shared knowledge one explicit home.
• Improved: Polished the Pinning article with a clearer opening on why agentic systems need pinned model, prompt, schema, and fixture choices, plus tighter prose and current pattern-marker wording.
• Sources: Added external source links across the Correctness, Testing, and Evolution section so readers can follow cited papers, books, standards, docs, and practitioner reports directly from the Sources blocks on every article in the section.
• Structural: Improved the Intent, Scope, and Decision-Making section index by adding Scope Creep to the list and relabeling it as entries rather than only patterns, so the index now covers patterns and antipatterns together.

Metrics

• Total articles: 257
• Coverage: 257 of 262 proposed concepts written (98.1%)
• Articles changed since last checkpoint: 2 new, 1 edit, 30 sources-linked

2026-05-08

What’s New

• Improved: Sharpened What Are Design Patterns? so the article names moves, not problems – a pattern names the solution shape that keeps resolving the same recurring tension, not the tension itself. Three sentences in the lead, the Alexander paragraph, and the “Patterns Are a Thinking Tool” opener tightened to match.
• Improved: Anchored the production technology by name in EACP’s introduction surfaces – Welcome, the orientation flag at Introduction, How This Book Writes Itself, the Meta Report opener, and the Colophon now read “the Bartley engine” on first mention rather than the older “agentic system” or “autonomous improvement engine” phrasings.
• Improved: Backfilled the body of the type-marker admonition on Delegation Chain so the orange “PATTERN” callout now carries the canonical one-line definition like every other entry.
• Improved: Updated the homepage description and ad-network blurbs to report the current corpus size at 250+ patterns (up from the older 190+ figure).
• Other: New cover with refreshed 3:4 portrait master, breathing room under the masthead, and the Bartley Editions publisher mark added to the chrome and the foot of the Colophon.
• Other: Two house themes added to the picker – Bartley Light (the new default) and Bartley Dark – alongside the five mdbook stock themes. Body text is set in Newsreader; the running head uses Cinzel; the palette mirrors bartleyeditions.com.
• Structural: What’s New, Pattern Map, and Meta Report now live as the last children of the Introduction section in the table of contents (where they have always belonged), restored from a brief stretch where they sat next to Cover and Colophon.

Metrics

• Total articles: 255
• Coverage: 255 of 256 proposed concepts written (99.6%)
• Articles changed since last checkpoint: 0 new, 4 edits

2026-05-03

What’s New

• New article: Plan-and-Execute – the agent architecture that splits the inner loop into a planner that thinks once, an executor that runs each step, and a re-planner that re-engages only when the plan needs to change; covers the three production variants (Vanilla, ReWOO, LLMCompiler) and when to pick this over ReAct or Plan Mode.
• New article: Pinning – the cross-cutting discipline of fixing model ids, prompts, schemas, and dependencies at immutable identifiers so your agent’s behavior doesn’t drift under you.
• Improved: Polished the Prompt Caching article with sharper jargon glosses (KV-cache, TTL), cleaner cost-section framing, and a tighter cache-invalidation warning; promoted from initial draft to edited.
• Improved: Updated the A2A article with corrected Linux Foundation governance attribution (A2A sits directly under the LF, not the Agentic AI Foundation), the accurate official-SDK list (five languages, not six – community Rust noted separately), the precise v1.0 release date (March 12, 2026), and the three cloud platforms now running A2A in production.
• Improved: Edited the Agentic Context Engineering article to tighten the explanation of how ACE relates to its parent pattern, sharpen the benchmark numbers, and split a long scenario for readability.
• Improved: Edited the Context Offloading article: corrected the seven-pattern list to its primary-source canonical names (Give Agents A Computer, Multi-Layer Action Space, Progressive Disclosure, Offload Context, Cache Context, Isolate Context, Evolve Context), credited Lance Martin by name for the LangChain framing, added canonical URLs to the Manus, Cursor, LangChain, and Anthropic citations, and tightened a couple of prose details.
• Sources: Added a Sources section to the Trust Boundary article, tracing the concept from Saltzer and Schroeder (1975) through Howard and LeBlanc’s Writing Secure Code (2003), Microsoft’s STRIDE framework, Shostack’s Threat Modeling (2014), and OWASP.
• Sources: Added a Sources section to Source of Truth, crediting Hunt and Thomas (DRY), Bill Inmon (data warehousing), and E. F. Codd (relational model) for the article’s intellectual lineage.
• Sources: Added a Sources section to the Consistency article crediting Jim Gray, Härder and Reuter, Eric Brewer, Gilbert and Lynch, and Werner Vogels for the transaction model, ACID, the CAP theorem, and eventual consistency.
• Sources: Added a Sources section to the Failure Mode article, crediting the FMEA tradition (US military 1949 and NASA Apollo), Charles Perrow’s Normal Accidents, Lamport et al.’s Byzantine Generals paper, and Werner Vogels’s “Everything Fails All the Time” alongside the Google SRE book.
• Sources: Added a Sources section to the Dependency article crediting Parnas 1972, Fowler 2004, Evans 2003, semver.org, and the dependency-hell folk concept.
• Sources: Added external links to the Sources sections of eight Design Heuristics & Smells articles (Premature Optimization, Best Current Practice, Code Smell, Jagged Frontier, KISS, Local Reasoning, Make Illegal States Unrepresentable, YAGNI), so every cited paper, talk, book, or essay now has a working URL.

Metrics

• Total articles: 255
• Articles changed since last deploy: 2 new, 4 edits, 13 sources-linked

2026-05-01

What’s New

• New article: Agentic Context Engineering – treat the agent’s working context as an evolving structured playbook updated incrementally by three roles (Generator, Reflector, Curator) instead of monolithic rewrites, to avoid the brevity-bias and context-collapse failure modes that destroy naive self-rewriting loops.
• New article: Context Offloading – the discipline of routing big tool outputs to the filesystem and giving the agent a short summary plus a file reference, so the active context stays focused on the work instead of drowning in tool exhaust.
• Improved: Polished the new Agentic Engineering article – tightened the prose, normalized “subagent” usage, and replaced two broken Open Library citations with correct works.
• Improved: Refreshed the Agent Teams article against Anthropic’s published Agent Teams documentation – corrected the coordination model (teams share one workspace coordinated by file locking on task claims, not separate worktrees), added coverage of plan-approval gating, task lifecycle hooks, and reusable subagent definitions, and tightened the Sources attribution to match Google ADK’s actual orchestration vocabulary.
• Improved: Polished the Structured Outputs article – sharper prose, a fixed typo, and a tighter Sources opener; promoted from initial draft to edited.
• Sources: Blast Radius now credits the people who developed the underlying ideas – Saltzer and Schroeder for least privilege (1975), Michael Nygard for the bulkhead pattern (Release It!), AWS for cell-based cloud architecture vocabulary, and Charity Majors for limited-blast-radius deployment as a discipline.
• Sources: Eight articles in the Computation and Interaction section now have hyperlinked sources – Turing’s “On Computable Numbers”, Knuth’s TAOCP, Hartmanis & Stearns 1965, Fielding’s REST dissertation, Dijkstra 1965, Hoare’s CSP, Hewitt’s actor formalism, Rob Pike’s “Concurrency Is Not Parallelism”, Cerf & Kahn 1974, Saltzer-Reed-Clark’s end-to-end argument, Berners-Lee’s WWW proposal, Anthropic’s MCP, Google’s A2A, and more all now resolve to the canonical primary source.
• Sources: Source citations across the Team Topologies cluster (Conway’s Law, Inverse Conway Maneuver, Stream-Aligned Team, Enabling Team, Platform as a Product, Thinnest Viable Platform, Team Cognitive Load, Ownership, Organizational Debt, Bounded Agency) are now hyperlinked to their canonical sources on the web – Skelton & Pais’s Team Topologies, Conway’s 1968 Datamation paper, Brooks’s Mythical Man-Month, Sweller’s cognitive-load research, the CNCF Platforms whitepaper, Bird and Greiler’s Microsoft code-ownership studies, the DORA 2025 report, and others.
• Sources: Added external links to every Sources entry across all eleven Security and Trust articles, so citations now jump straight to the canonical paper, book, or post.

Metrics

• Total articles: 253
• Articles changed since last deploy: 2 new, 3 edits, 21 sources-linked

2026-04-27

What’s New

• New article: Prompt Caching – pin the unchanging part of your prompt at the front so the provider can reuse its computed state and bill the repeat at a fraction of the cost.
• New article: Compound Engineering – make every shipped lesson land on a durable, agent-readable surface (instruction file, skill, hook, subagent, test) before the work closes, so the next feature is genuinely cheaper than the last.
• New article: Agentic Engineering – the professional discipline of orchestrating coding agents, supervising their work, and reviewing the output, where humans now write less than 1% of the code directly.
• New article: Structured Outputs – how to constrain a model’s response to a known schema so the next program in the pipeline can parse it without guessing.
• New article: Agent Registry – the directory of identities, capabilities, and ownership for every agent in the fleet, the substrate that has to ship before any policy you’d want to bind to “which agent is this” can work.
• New article: LLM-as-Judge – how to use one model to score another’s output against a rubric, the practical workhorse of agentic evaluation, with the four canonical biases (position, verbosity, self-preference, authority) and the de-biasing playbook for each.
• New article: Agent Gateway – the runtime control plane that brokers every tool call between agents and tools, centralizing authentication, authorization, audit, and policy enforcement so credentials describe potential and gateway policy describes what’s allowed right now.
• New article: Runtime Governance – the discipline of moving every policy decision onto the action path itself, where each tool call is ruled allow, throttle, sandbox, escalate, or block at machine speed before it reaches the world.
• Improved: MCP article corrected – dropped the spurious “MCP v2.1” version label (no such version exists; the latest spec is 2025-11-25), reframed Server Cards as the still-Draft SEP-1649 proposal it actually is, and re-attributed the 2026 MCP roadmap from the “AAIF technical steering committee” to its actual author, lead maintainer David Soria Parra.
• Improved: Model refreshed for the hybrid-reasoning era – the “Models differ” passage now leads with hybrid models as the dominant 2026 frontier shape (GPT-5, Claude Opus 4.5, Gemini 2.5 each named with their effort/router/thinkingBudget knob), the multimodal claim is hedged so it tracks reality (text+images universal; audio+video vendor-dependent), and the Stochasticity force now acknowledges the batch-invariance engineering recipe that can deliver bit-identical output even though most production APIs don’t enable it.
• Improved: Triple Debt Model and Vella’s Middle Loop vocabulary now land across four articles – the Technical Debt antipattern names intent debt as a fourth, artifact-level variant alongside cognitive and agentic debt, and Human in the Loop, Steering Loop, and Harness Engineering each name supervisory engineering, Annie Vella’s empirical decomposition of middle-loop work into directing, evaluating, and correcting, anchored on her 158-engineer / 28-country longitudinal study.
• Improved: Organizational Debt now renders a Related Patterns table and shows up as a connected node in the local-graph widget – 10 typed links to Conway’s Law, Ownership, Team Cognitive Load, Inverse Conway Maneuver, Stream-Aligned Team, Enabling Team, Platform as a Product, Bounded Agency, Technical Debt, and Agent Registry let readers traverse the socio-technical neighborhood in both directions for the first time.
• Improved: Tightened the prose in Runtime Governance and added direct links to the primary sources behind it (Oracle, Microsoft, Microsoft Open Source, Prefactor, and the arXiv preprint).
• Improved: Polished the Compound Engineering article – eliminated em-dash overuse in the prose and tightened a couple of awkward phrasings, promoting it from initial draft to edited.
• Improved: Polished the LLM-as-Judge article – tightened sentence rhythm, trimmed prose em-dashes, and promoted the entry from initial draft to edited.
• Improved: Corrected the Skill article’s Sources section to list Anthropic’s actual launch partners (Box, Canva, Notion, Rakuten) and a current cross-vendor adopter list spanning major coding agents, data tools, application frameworks, and IDE plugins.
• Improved: The Smoke Test article got a prose polish – tighter Solution opener, a triple-conjunction run-on rewritten, a minor self-repetition fixed, and the “tier” framing clarified in two cross-reference lines.
• Improved: Corrected publication dates and benchmark figures in the Code Mode article’s Sources section, and added the March 2026 Cloudflare update that shipped Code Mode into MCP server portals by default.
• Improved: Polished the Agent Trace article – tighter sentence rhythm and reduced AI-style cadences while preserving every example, source, and link.
• Improved: Edited the Agent Registry article for prose quality – tighter sentences, broken-up parallel structures, and a clearer Solution section explaining why a registry has to ship before any policy that would bind to it.
• Improved: The Feedback Sensor article was edited – linked the inferential-sensor example to the new LLM-as-Judge entry, replaced an unattributed “studies show” claim with a sharper one-liner, and tightened the closing paragraph on infrastructure cost.
• Sources: Added a Sources section to the Event article – credits the structured-design community for the broad pattern, Hohpe and Woolf’s Enterprise Integration Patterns for the messaging vocabulary, Martin Fowler’s 2017 taxonomy article for the four senses of “event-driven,” and Greg Young for CQRS and event sourcing.
• Sources: The Transaction article now credits its intellectual roots – Jim Gray’s 1981 transaction-concept paper, Härder and Reuter’s 1983 paper that coined the ACID acronym, the 1992 Gray-Reuter classic, Martin Kleppmann’s Designing Data-Intensive Applications, and Garcia-Molina and Salem’s 1987 Sagas paper – with direct links to each primary source.
• Sources: The Module article now credits its intellectual roots – Parnas’s 1972 information-hiding paper, Yourdon and Constantine’s Structured Design (1979) for cohesion, Ousterhout’s A Philosophy of Software Design for the deep-modules framing, and Wirth’s 1971 stepwise-refinement paper – with direct links to each primary source.
• Sources: Added intellectual lineage to the Boundary article – credits Parnas’s 1972 information-hiding paper for the rate-of-change criterion, Evans’s Domain-Driven Design for bounded-context-driven boundary placement, and Nygard’s Release It! for the failure-containment view of boundaries.
• Sources: Sources sections in Affordance, User Feedback, and UX now link directly to canonical references (Gibson, Norman, Nielsen, Gaver, Myers, Raymond, Walker NYT) so you can jump from citation to primary source in one click.
• Sources: Sources for four product-judgment articles – Bottleneck, Crossing the Chasm, Product-Market Fit, and Zero to One – now link to the original works and posts they cite (Goldratt’s The Goal, Rogers’s Diffusion of Innovations, Moore’s Crossing the Chasm, Andreessen’s PMarca essays, Sean Ellis’s Startup Pyramid, Blank’s Four Steps to the Epiphany, Thiel and Masters’s Zero to One, and Blake Masters’s CS183 class notes).
• Structural: Connections graph fills out further – 70 new typed back-links across 29 articles in 8 sections (so every relation involving a socio-technical-systems pattern reciprocates correctly from the target side) plus another 40 new typed back-links across 17 structure-and-decomposition articles, and two silent dead edges in the Agent Gateway and Organizational Debt articles are now live links rather than missing nodes.

Metrics

• Total articles: 251
• Articles changed since last deploy: 8 new, 13 edits, 12 sources-linked

2026-04-26

What’s New

• New article: Smoke Test – the cheapest, fastest test you can run, with three scenarios (classical CI, post-deploy automatic-rollback, and agentic verification loops) and a five-question checklist for designing one.
• New article: Agent Trace – how to capture each agent run as a tree of spans (model calls, tool calls, sub-agent dispatches) so you can debug a wrong answer, attribute its cost, correlate sub-agents back to the parent, and replay the run against a new model.
• Improved: The Artifact article was edited so its examples and closing argument no longer echo Externalized State – the migration scenario was replaced with an SRE incident-handoff scenario, and the Consequences section was rewritten to break stacked tricolons.
• Improved: The Permission Classifier article got a prose pass – tighter paragraphs and a much lighter em-dash count.
• Improved: The Interactive Explanations article was polished – the Consequences “costs” paragraph was rewritten for cleaner rhythm, and the Sources passage on cognitive debt was tightened.
• Sources: Added a Sources section to the Affordance article crediting J.J. Gibson (who coined the term in 1966 and developed it in 1979), Don Norman (who imported it into design in 1988 and refined it with the affordance/signifier distinction in 2013), and William Gaver (whose 1991 CHI paper brought it into HCI).
• Sources: Added a Sources section to the CRUD article crediting James Martin (who coined the acronym in Managing the Data-base Environment, 1983), Chamberlin and Boyce (whose 1974/1976 SEQUEL papers gave us the SQL DML verbs CRUD abstracts), and noting that the familiar HTTP-verb mapping is a community convention – not, as commonly believed, a prescription from Roy Fielding’s REST dissertation.
• Sources: Added a Sources section to the Continuous Delivery article crediting Jez Humble and David Farley (whose 2010 Continuous Delivery book named the deployment pipeline and won the 2011 Jolt Excellence Award), the earlier Agile 2006 ThoughtWorks paper by Dan North, Chris Read, and Jez Humble that introduced the deployment-pipeline concept, and Forsgren, Humble, and Kim’s 2018 Accelerate – which provided the empirical case for CD’s effect on performance through the DORA metrics.
• Structural: Reworked the Connections map at the top of every article. The widget is now an adaptive square that grows with the page width, and the related concepts spread out across concentric rings instead of crowding into a single fixed-radius cluster. Labels no longer overlap, and edges are pushed apart so two related concepts never line up as a single ray through the hub.
• Structural: The Related Patterns section at the bottom of every article is now a sortable table – Relation, Article, and Note columns, with clickable Relation and Article headers that re-sort the list and toggle ascending/descending. The default order matches the old grouped-by-relation bullet list.

Metrics

• Total articles: 245
• Articles changed since last deploy: 2 new, 3 edits, 3 sources, 1 structural

2026-04-25

What’s New

• New article: Agentic Manual Testing – how to hand an agent a short charter plus a browser driver and let it do the integration-QA clicking, typing, and watching that a human tester used to do before every release.
• New article: Artifact – a durable, named, inspectable product of work; the missing vocabulary entry for a word the book has been using everywhere without defining.
• New article: Permission Classifier – the third path between approving every agent action by hand and running unattended with no safety check, where a small classifier model judges each proposed action in real time and routes it to auto-approve, escalate, or block.
• New article: Interactive Explanations – the trick of asking the agent that just wrote your code to build an animated, scrubbable visualization of the algorithm, so you form intuition against live execution instead of a paraphrase.
• Improved: The RAG Poisoning article gained an OWASP LLM08:2025 reference and a fifth practical defense – permission-aware vector databases for multi-tenant embedding leakage.
• Improved: The Technical Debt article was expanded to cover the three AI-era debt variants – cognitive debt (the gap between code shipped and code understood), comprehension debt (teams merging more code than anyone reads), and agentic debt (the hidden infrastructure cost of running agents without guardrails) – with attributions to Storey, Osmani, the New Stack, JetBrains, and Gartner.
• Improved: The DWIM article was edited for prose quality and consistent dash rendering.
• Improved: The Greenfield and Brownfield article was edited for prose quality and consistent dash rendering.
• Improved: The Regenerative Software article was edited for prose quality and consistent dash rendering.
• Improved: The Agentic Manual Testing article was edited for prose quality – tighter rhythm in the Problem section, a fresh third scenario element in the Solution disciplines, and a cleaner em-dash budget overall.
• Sources: Added external links to every primary reference in four Intent and Scope articles – Architecture Decision Record, Design Doc, Spec-Driven Development, and Specification – so readers can jump straight to the originating papers, essays, and books.
• Structural: Rebuilt the Cover page’s Browse section with per-section descriptor paragraphs – each of the 14 sections now gets a one-line purpose, a handful of representative entries, and a link to see them all – instead of the old 245-bullet flat list.

Metrics

• Total articles: 249
• Articles changed since last deploy: 4 new, 7 edits, 1 structural

2026-04-24

What’s New

• New article: Ship – the root verb the rest of the book leans on, with a precise definition that preserves the classical meaning (in users’ hands, in a version you can no longer silently change) and names the three agentic-era shifts in who carries the work, what counts as shippable, and at what cadence.
• New article: Footgun – the design-property lens for features, tools, and defaults that are easy to use wrong and hard to use right, with a four-move mitigation taxonomy and an agent-tool audit you can run today.
• New article: DWIM – the “Do What I Mean” principle, traced from INTERLISP’s 1970s typo-correcting helper through modern agentic harnesses that infer intent from sloppy input.
• New article: Greenfield and Brownfield – how to name the mode of work so the agent applies the right patterns (and stops adding backwards-compatibility code to a clean-slate project).
• New article: Regenerative Software – treating code as a disposable output of durable specs, boundaries, and evals, so individual components can be deleted and regenerated on a cadence instead of maintained in place.
• Improved: The Ship article was edited for prose quality – trimmed scaffolding, reworked the em-dash budget down to zero in-prose, and cut filler phrases. Meaning and structure preserved.
• Improved: The Bounded Agency article was polished – tighter sentences and clearer attribution in the organizational-scenario narratives.
• Improved: The Tool Sprawl article was revised for voice and cadence, varying the three example scenarios so they no longer read as stamped from the same template.
• Improved: The REPL article was revised – unified the term to the canonical “read-eval-print loop,” split an over-long walkthrough paragraph into two scannable ones, and cleaned up typographic details for consistency with the rest of the section.
• Improved: Added direct links to cited works in the Cascade Failure and Continuous Integration articles – every reference now goes straight to the primary source.
• Structural: Strengthened navigation in the Security and Trust section – 13 foundational articles now link back to the newer concepts (Adversarial Cloaking, Agent Trap, Agentic Payments, RAG Poisoning) that depend on them.

Metrics

• Total articles: 248
• Articles changed since last deploy: 5 new, 5 edits, 1 structural

2026-04-23

What’s New

• New article: Load-Bearing – the structural-engineering term for code, comments, tests, or instructions whose removal breaks things in non-obvious ways, with a specific focus on why agents are uniquely dangerous around them.
• New article: Sweep – the discipline of applying one rule uniformly across many files, with a decision rule for picking between regex, codemod, and agentic execution modes, and the safety practices that keep the blast radius manageable.
• New article: Agent-Computer Interface (ACI) – the discipline of designing tools, affordances, and interaction formats for language-model agents rather than humans, grounded in the Princeton SWE-agent result that moved a coding agent from near-zero to state-of-the-art on SWE-bench by changing only the command surface.
• New article: Tool Sprawl – the antipattern where an agent’s tool catalog grows past the model’s ability to select cleanly among its members, with accuracy collapsing as capabilities keep being added.
• New article: REPL – the read-evaluate-print-loop shape that Claude Code, Aider, and most coding agents inhabit, traced from Lisp’s 1960s origin through today’s agentic harnesses.
• New article: Bounded Agency – the organizational authority envelope that makes delegation to humans and AI agents governable, named by Matthew Skelton at QCon London 2026 and anchored in the OWASP LLM06 Excessive Agency failure mode.
• Improved: The Load-Bearing article was edited for prose polish – added an inline gloss for Chesterton’s Fence on first use and aligned the Related Patterns bullet separators with section convention.
• Improved: The Sweep article was edited for prose polish – tightened cadence with natural contractions and aligned its bullet separators with the Encyclopedia’s em dash convention.
• Improved: The Agent-Computer Interface (ACI) article was edited for prose polish – fixed the concept-template header order, added an inline gloss for tool sprawl on first mention, and tightened a sentence in the naive-search example.
• Improved: The Sandbox article was updated with 2026’s agentic-sandboxing landscape – microVM isolation (Docker Sandboxes, E2B, Cloudflare Sandbox) as a distinct mechanism between containers and full VMs; Claude Code’s OS-level enforcement through bubblewrap on Linux and Seatbelt on macOS; and a new Consequences section on reasoning-agent bypasses, drawing on Ona’s study of Claude Code escaping its own denylist and sandbox.
• Structural: The cover page now carries four deploy-refreshed “brag cards” above the table of contents – total Articles, plus Patterns / Antipatterns / Concepts breakdowns – refreshed on every deploy by a new sync-cover-counts script. Stale count claims removed from the intro paragraph.

Metrics

• Total articles: 243
• Articles changed since last deploy: 6 new, 4 edits, 1 structural

2026-04-19

What’s New

• New article: Brief – the short, frame-setting document that names what you’re building and why, before any spec exists.
• New article: Domain-Oriented Observability – instrument business-meaningful events (cart abandoned, payment declined, order placed) as first-class telemetry, so dashboards track outcomes and not just process health.
• Improved: The Brief article was edited for prose quality – tightened the opening, made the “what matters most” example concrete, and rewrote the counter-example paragraph to read less mechanically.
• Improved: The Least Privilege article was refreshed with the modern agent-security vocabulary: excessive agency (the AWS/OWASP-named risk), permission boundary (the policy ceiling), and agent gateway (runtime tool-call enforcement), plus concrete citations to the MCP spec, AWS Well-Architected, and OWASP Top 10 for LLMs.
• Improved: The Memory article was updated with Claude Code’s Auto Memory as a concrete example of automated memory extraction, and the Sources bullet was expanded to cover both CLAUDE.md and the new MEMORY.md layer.
• Improved: The Domain-Oriented Observability article was edited for tighter prose and clearer framing.
• Sources: Expanded the Threat Model article’s Sources with the 2020 Threat Modeling Manifesto and the current agent-specific references (OWASP Top 10 for Agentic Applications and MITRE ATLAS).
• Sources: Improved the Test Oracle article with a Sources section crediting the originators of test-oracle terminology, the oracle problem, property-based testing, and the standard modern survey.
• Sources: Added a Sources section to the Invariant article, crediting Hoare, Dijkstra, Meyer, and Evans.
• Structural: Added reciprocal navigation links so readers of Silent Failure, Failure Mode, Invariant, and Test Oracle can see at a glance that Fail Fast and Loud is the directly related pattern.
• Structural: Added canonical external links to every citation in the Structure and Decomposition chapter’s Sources sections – readers can now jump directly from an acknowledgment to the originating paper, essay, or book.
• Other: Updated the cover to include the three new Agentic Software Construction articles shipped in the previous round (Task Horizon, Deep Agents, Reflexion).

Metrics

• Total articles: 237
• Articles changed since last deploy: 2 new, 4 edits, 3 sources, 2 structural

2026-04-18

What’s New

• New article: Task Horizon – the duration an agent can work coherently on its own, the scoping concept every long-running run is implicitly negotiating with.
• New article: Deep Agents – the four-pillar recipe (explicit planning, sub-agent delegation, persistent memory, extreme context engineering) that turns a shallow loop into a harness capable of multi-hour tasks, and why Claude Code, Codex, and LangChain’s deepagents SDK all end up shaped the same way.
• New article: Reflexion – how to turn failed attempts into smarter retries by forcing the agent to articulate what went wrong before trying again.
• Improved: The A2A article was refreshed to match the v1.0 specification – Signed Agent Cards for identity verification at discovery time, gRPC as a peer binding alongside JSON-RPC, three task-delivery modes (polling, streaming, webhooks), multi-tenant endpoints, calibrated SDK coverage, and the Agent Payments Protocol as the first real-world A2A extension.
• Improved: Added Fowler/Garg/Morris vocabulary (Knowledge Priming, Encoding Team Standards, Context Anchoring, Design-First Collaboration, in/on/out of the loop) as synonyms across five articles so readers arriving with those terms find our treatment – and corrected an attribution error in Steering Loop that credited the wrong authors with the inner/middle/outer loop model.
• Improved: The ReAct article was polished – tighter rhythm in the Consequences section, cleaner parenthetical asides, and a fixed stray formatting mark at the end of the file.
• Improved: The Orchestrator-Workers article got tighter prose, a fixed worker-count inconsistency, and long paragraphs broken into easier reads.
• Improved: The Task Horizon article was edited for sharper prose – tighter opening, cleaner transitions, and a better tie-in to Model Routing.
• Improved: The Deep Agents article was edited for sharper prose – split the long narrative paragraph into scene beats and tightened the Consequences into cleaner, more scannable pairs.
• Improved: The Reflexion article was polished with tighter prose and cleaner rhythm.
• Sources: Added a Sources section to the Worktree Isolation article, crediting the Git contributors who built git worktree and the AI-coding community that popularized running one agent per worktree.

Metrics

• Total articles: 235
• Articles changed since last deploy: 3 new, 7 edits, 1 sources audit

2026-04-17 (evening)

What’s New

• New article: ReAct – the thought-action-observation loop that turns a model into an agent, and the inner primitive that Plan Mode, Verification Loop, and Ralph Wiggum Loop all wrap.
• New article: Orchestrator-Workers – the pattern where a central agent decides what subtasks a goal requires on the fly, dispatches workers to handle each, and stitches the results back together.
• Improved: The MCP article now covers MCP v2.1 Server Cards (pre-connection capability discovery at /.well-known/mcp/server-card.json), notes Tasks as still experimental, and adds the 2026 MCP roadmap’s four priority areas.
• Improved: The Back-Pressure (Agent) article was polished for prose quality – removed an approximated epigraph and tightened the Sources section.
• Improved: The Fail Fast and Loud article was polished for prose quality, differentiating its examples from the paired Silent Failure antipattern.
• Improved: The Consumer-Driven Contract Testing article was polished for prose quality and fixed a misleading Related Patterns link.

Metrics

• Total articles: 232
• Articles changed since last deploy: 2 new, 4 edits

2026-04-17 (afternoon)

What’s New

• New article: Harness Engineering – the discipline of configuring the surfaces around a coding agent (instructions, tools, MCP, skills, sub-agents, hooks, approval policy, memory, compaction, back-pressure, isolation) so a fixed model reliably does work on your codebase.
• New article: Exploratory Testing – how to run charter-driven sessions that find bugs scripted tests were never written to catch, including AI-pair-testing for code built by agents.
• New article: Jagged Frontier – why AI capability is uneven in ways that don’t match human intuition about task difficulty, and why that shape is the reason verification loops, evals, and bounded-autonomy policies exist.
• New article: Back-Pressure (Agent) – how to pace an agent so it doesn’t overwhelm itself, its tools, or the humans around it, with concrete mechanisms for the rate-related failure modes that approval policies and autonomy scopes don’t catch.
• New article: Fail Fast and Loud – detect invalid state at its source and surface it in a way that’s impossible to ignore, so nothing builds on a broken foundation.
• New article: Consumer-Driven Contract Testing – let each consumer declare the parts of an API it depends on and verify every consumer’s contract before release, so changes that break a real caller never reach production.
• Improved: The Harness Engineering article was polished for prose quality – smoother rhythm, natural contractions throughout, a tighter Solution opener.
• Improved: The Exploratory Testing article was polished – fixed a voice inconsistency in the oracle-discipline bullet and tightened the opening of the Solution section so the advice reads in shorter, punchier sentences.
• Improved: The Jagged Frontier article was polished – the explanation of why the frontier has spikes and bays now has its own section, separate from the practical recognition heuristics.
• Improved: The Dark Factory article was polished for prose quality – tightened a loose claim, activated a passive sentence, and softened the register with natural contractions.
• Improved: The Progressive Disclosure article got tighter prose and a stronger closing.
• Improved: The Model article sharpened its treatment of LLM stochasticity – temperature zero reduces variance but does not guarantee bit-for-bit reproducibility, and the article now explains why (GPU math, tie-breaks, batched serving) with a source.
• Sources: Added a Sources section to the Instruction File article, crediting CLAUDE.md, .cursorrules, GitHub Copilot instructions, and the AGENTS.md open format.

Metrics

• Total articles: 227
• Articles changed since last deploy: 6 new, 6 edits, 1 sources audit

2026-04-17

What’s New

• New article: Dark Factory – the operating model where coding agents write, test, and ship production software while humans work only at the specification and governance layer, with an honest account of the preconditions, risks, and accountability questions.
• New article: Progressive Disclosure – the design principle of loading instructions, tool definitions, and reference material into an agent’s working memory only when they become relevant, organized as three tiers. The practical counter to Context Rot.
• New article: AgentOps – the operational discipline of monitoring, costing, and governing AI agents running in production.
• New article: Agentic Payments – how to let an autonomous agent pay for things without handing it authority it can misuse.
• Improved: The Skill article now reflects the Agent Skills open standard’s December 2025 cross-vendor launch – Microsoft, OpenAI, GitHub, Cursor, Figma, and Atlassian adopted the spec, so a skill file is now portable across major agentic coding harnesses.
• Improved: The AgentOps article was polished with sharper vignettes and tighter prose.
• Improved: The Agentic Payments article was polished with clearer examples and tighter prose.
• Improved: The Tool article’s Consequences section was tightened for sharper, more concrete prose.
• Improved: The Code Mode article got a tighter Problem framing and a cleaner sandbox-liability note.
• Improved: The Threat Model article was polished for tighter prose.
• Sources: Added intellectual lineage to the Approval Policy article – Saltzer & Schroeder’s 1975 security principles, Anthropic’s Claude Code permission model, and the Knight Institute’s levels-of-autonomy framework.
• Sources: Added intellectual lineage to the Idempotency article – from Benjamin Peirce’s 1870 mathematical coinage through the HTTP RFCs and Stripe’s idempotency-key pattern to the current IETF draft.
• Infrastructure: Every article now lives at a flat, slug-based URL (e.g. /dark-factory instead of /agent_governance_and_feedback/dark_factory.html). Old hierarchical URLs redirect automatically. Slugs are permanent – once an article ships, its URL never changes, even if the article moves between sections. Inbound links keep working.
• Fix: The local-graph widget (the “Nearby Patterns” chart on every article page) was missing on many pages; it is now restored everywhere.
• Fix: The /introduction landing URL no longer loops back on itself; it resolves cleanly on the first click.

Metrics

• Total articles: 223
• Articles changed since last deploy: 4 new, 6 edits, 2 sources audits, plus a site-wide URL scheme migration and two navigation fixes

2026-04-15 (afternoon)

What’s New

• New article: Spec-Driven Development – the workflow that forms around a written spec and keeps agents aligned as systems grow, at three rigor levels from spec-first to spec-as-source.
• New article: Code Mode – give the agent a small API and a sandbox, and let it write code that calls tools instead of emitting JSON one step at a time.
• Improved: The Spec-Driven Development article was polished – the four core disciplines are now a clearer bullet list, the subtitle is cleaner, and the Feynman epigraph is sourced.
• Improved: The Eval article’s Sources now reflect the 2026 SWE-bench landscape – SWE-bench Verified (the de-facto scoreboard) and SWE-bench Pro (its harder successor).
• Improved: The Harnessability article’s Further Reading gained two 2026 companion pieces – OpenAI’s Codex App Server follow-up and HumanLayer’s harness engineering deep dive.
• Improved: The Test Pyramid article was tightened – corrected attribution for “The Practical Test Pyramid” to Ham Vocke, fixed a duplicate link in Further Reading, and softened an uncited claim.
• Improved: The YAGNI article got a clearer name for the “familiar-shape bias” force and a sharper closing paragraph on the optionality YAGNI preserves.
• Improved: The Approval Policy article got a punchier Consequences section and cleaner cross-reference formatting.
• Sources: Tightened intellectual lineage on two articles: Feedforward (corrected Harold S. Black’s patent history, dated Marshall Goldsmith’s essay, named the I. A. Richards lecture) and Smell (Code Smell) (added the origin WikiWikiWeb entry and Fowler’s bliki definition).
• Meta: The thirty-first meta report sharpened the write-vs-edit balance to unblock a persistent write under-firing pattern. See the Meta Report.

Metrics

• Total articles: 212
• Articles changed since last deploy: 11 reader-visible cycles (2 new articles, 6 edits, 2 sources audits, 1 meta report)

2026-04-15

What’s New

• New article: Test Pyramid – how to allocate testing effort across fast unit tests at the base and a small number of end-to-end tests at the top, with an agentic variant that reorganizes the layers by uncertainty tolerance.
• Improved: The Code Smell article was tightened with a crisper opening, sharper voice, and an explicit nod to Fowler’s canonical catalog.
• Improved: The Adversarial Cloaking article gained sharper prose and a fourth danger property – persistence by default – linking the threat to decades of search-engine cloaking.
• Improved: The Model Routing article was refreshed for 2026 with cascade routing as a distinct fifth form, named production router infrastructure (RouteLLM, LiteLLM, Bifrost), GPT-5 internal routing as an industry signal, and new benchmark sources (RouterBench, RouterEval).
• Structural: Tightened cross-linking across the Correctness, Testing & Evolution section so Related Patterns now resolve in both directions – 85 new reciprocal links help readers jump between connected ideas.
• Sources: Added or tightened intellectual lineage on four articles: Service Level Objective (Google SRE book, the SRE Workbook, Alex Hidalgo’s Implementing SLOs), Test-Driven Development (TDAD research paper referenced by arXiv ID and benchmark), UX (Don Norman, Jakob Nielsen, 2003 Steve Jobs profile), and Big Ball of Mud (Ward Cunningham’s original 1992 technical debt metaphor).
• Meta: The thirty-first meta report rebalanced write and sources coefficients after sources dominated recent cycles. See the Meta Report.

Metrics

• Total articles: 229
• Coverage: 229 of 235 proposed concepts written (97%)
• Articles changed since last deploy: 11 content cycles (1 new article, 3 edits, 1 groom, 4 sources audits, 1 meta, 1 no-op critique)

2026-04-12 (night)

What’s New

• New article: Code Review – how having fresh eyes on every change catches what tests and the author’s own familiarity miss, and why agent-generated code makes review more important, not less.
• New article: Adversarial Cloaking – how attackers detect AI agent visitors and serve them poisoned web pages invisible to human reviewers.
• Improved: The Context Window article now reflects the 2026 token landscape (128K to 10M) with a note on attention degradation at million-token scale.
• Improved: The Handoff article gained sharper prose, a concrete Stripe API migration scenario, and the “context dump fallacy” concept.
• Improved: The Agent article gained the agentic-vs-vibe-coding distinction and 2026 production adoption data from Gartner and LangChain.
• Improved: The Code Review, Human in the Loop, and Side Effect articles were polished for prose quality.
• Improved: The Prompt Injection article gained Simon Willison’s Lethal Trifecta risk framework: three conditions that determine when injection becomes critically dangerous.
• Sources: Added intellectual lineage to four articles: Threat Model (Shostack, STRIDE, SDL), Observability (Kalman, Twitter, Honeycomb, Sridharan, Google Dapper), Sandbox (Bill Joy’s chroot, FreeBSD Jails, Java sandbox, seccomp), and Bottleneck (Goldratt’s Theory of Constraints, Thomas Reid).
• Structural: Prerequisite links across all articles now form a verified directed acyclic graph – following “Understand This First” links always leads to foundation concepts, never in circles.
• Meta: Two meta reports confirmed the engine in stable equilibrium: sources coverage passed 50%, pipeline steady at 4, no parameter changes needed. See the Meta Report.

Metrics

• Total articles: 217
• Articles changed since last deploy: 20 cycles (2 new articles, 7 edits, 4 sources audits, 2 meta reports, 3 research rounds, 1 sweep, 1 no-op critique)

2026-04-12 (evening)

What’s New

• New article: Feedback Flywheel – the cross-session retrospective loop that turns repeated corrections into permanent instruction-file rules, with first-pass acceptance rate as the metric.
• New article: Organizational Debt – the accumulated cost of structural shortcuts in how teams are organized and decisions are made, compounding silently until the organization can’t move.
• New article: Delegation Chain – the path authority follows from a human through agents, where each link can amplify or misdirect the original intent. Covers the confused deputy problem and practical chain-of-custody tracking.
• New article: Cascade Failure – when one component’s failure triggers failures in others, creating a chain reaction that can bring down an entire system faster than anyone can respond.
• New article: Handoff – how to transfer context, authority, and state between agents or sessions without losing what matters or carrying along what doesn’t.
• Improved: The Domain Model article gained tighter prose and cleaner source attributions.
• Improved: The Feedback Flywheel, Delegation Chain, Crossing the Chasm, Organizational Debt, and Cascade Failure articles were all polished for prose quality – tighter sentences, varied structure, and better paragraph rhythm.
• Improved: Every page now has a single H1 heading tag for better SEO clarity, and all major sections now group their patterns under thematic sub-headings for easier scanning.
• Sources: Added intellectual lineage to Concurrency (Dijkstra, Hoare, Hewitt, Pike, async/await) and Human in the Loop (Wiener, Bainbridge, Shneiderman).
• Meta: The engine confirmed its stochastic self-correction model (write recovered from zero to four articles in one period), validated the zero-pressure fix and sources gate, and bumped research priority to replenish a draining proposal pipeline. See the Meta Report.

Metrics

• Total articles: 215
• Articles changed since last deploy: 20 cycles (5 new articles, 6 edits, 2 sources audits, 1 groom audit, 1 meta report, 2 research rounds, 2 sweeps, 1 no-op critique)

2026-04-12

What’s New

• New article: Context Rot – why agent output quietly degrades as inputs grow, even when the context window is nowhere near full, and what the existing context patterns are really fighting.
• Improved: The Agent Sprawl antipattern gained tighter prose so the governance argument lands harder.
• Improved: The Question Generation article was polished with tighter sentences and a reworked Sources section citing the intellectual lineage (Gause & Weinberg 1989, Brooks 1986).
• Improved: The Context Rot article was edited for prose quality: varied rhythm, a fixed internal contradiction, and natural contractions throughout.
• Sources: Five articles gained intellectual-lineage credits: User Feedback (Don Norman, Jakob Nielsen, Brad Myers, Unix rule of silence), KISS (Kelly Johnson, Hoare, Dijkstra, UNIX tradition, Rich Hickey), Premature Optimization (the full Knuth attribution story, Bentley, Gregg, Fowler), Coupling (Stevens, Myers, Constantine, Yourdon, Parnas), and Tool Poisoning (Invariant Labs, CyberArk, Elastic Security Labs, Simon Willison).
• Structural: Added 62 missing reciprocal Related Patterns backlinks across two chapters: Operations and Change Management (29 links) and Product Judgment (33 links).
• Meta: Two meta reports this period. The twenty-fifth confirmed the em-dash gate held for a second straight period and caught a competitor-name leak in a Sources section mid-cycle, fixing the write procedure on the spot. The twenty-sixth implemented a zero-pressure exclusion fix so the stochastic sampler no longer wastes cycles on actions with nothing to do. See the Meta Report.

Metrics

• Total articles: 210
• Articles changed since last deploy: 21 cycles (1 new article, 3 edits, 5 sources audits, 2 groom audits, 2 meta reports, 2 research rounds, 2 critique no-ops, 2 sweep no-ops, 2 freshness checks)

2026-04-11 (evening)

What’s New

• New article: Business Capability – how to name what a business does (independent of teams, processes, and technology) so that strategy, software, and agent tasks all share the same stable anchors.
• New article: Parallel Change – change an interface safely by adding the new form, migrating callers at their own pace, and removing the old form last.
• New article: Deprecation – how to retire a feature, endpoint, or field on a schedule so callers have fair warning and you have evidence the removal is safe.
• New article: Evolutionary Modernization – how to modernize legacy systems as an ongoing engineering practice instead of a risky big-bang rewrite.
• New article: Agent Sprawl – the antipattern of autonomous agents proliferating across an organization faster than governance can track them, and how to escape it by treating the agent fleet as production infrastructure.
• New article: Question Generation – a pattern for making the agent interview you (in categories, one at a time, with default answers) before it writes any code.
• Improved: The AI Smell article gained a new team-dynamics smell – shipping agent-written code onto a reviewer without first reviewing it yourself – and a tighter discussion of authorship ownership.
• Improved: The Thread-per-Task article gained a concrete before-and-after transcript showing how context scrolling degrades agent output and how a fresh thread restores quality.
• Improved: The Deprecation, Parallel Change, Service Level Objective, Business Capability, and Evolutionary Modernization articles were all polished for prose quality – smoother rhythm, tighter sentences, and cleaner cross-reference lists.
• Sources: Added intellectual lineage to two articles: Least Privilege (Saltzer and Schroeder’s 1975 paper where the principle was coined, plus Mark Miller’s object-capability work) and Thread-per-Task (Drew Breunig, the Manus team, and Anthropic on context-window degradation).
• Structural: Added 48 missing reciprocal Related Patterns backlinks across the Intent and Scope (21 links) and Agent Governance and Feedback (27 links) chapters, so every relationship is now visible from both sides.
• Meta: Two meta reports this period. The first upgraded the write procedure’s em-dash check to a hard blocking gate after four drafts shipped over budget; the second confirmed the fix worked (new drafts shipped at 0 and 1 em dashes) and confirmed the recent sources starvation was ordinary variance, not a bug. See the Meta Report.

Metrics

• Total articles: 209
• Articles changed since last deploy: 24 cycles (6 new articles, 7 edits, 2 sources audits, 2 groom audits, 2 meta reports, plus research and no-op cycles)

2026-04-11

What’s New

• New article: Thinnest Viable Platform – build the smallest internal platform that lets teams deliver autonomously, then grow it only when demand is real.
• New article: Service Level Objective – how to pick a reliability target, measure how often you meet it, and spend the slack as an error budget that governs when to ship and when to slow down.
• New article: Parallel Change – change an interface safely by adding the new form, migrating callers at their own pace, and removing the old form last.
• New article: Business Capability – how to name what a business does (independent of teams, processes, and technology) so that strategy, software, and agent tasks all share the same stable anchors.
• Improved: The Strangler Fig article gained tighter prose, better paragraph rhythm in the Consequences section, and a new cross-reference to Decomposition.
• Improved: The Thinnest Viable Platform article was edited for tighter prose and more concrete framing of the agent governance example.
• Structural: Renamed the Feedback article to User Feedback to clearly distinguish it from Feedback Loop; updated all cross-references.
• Meta: The engine cut the sources coefficient last cycle and discovered sources fired zero times in ten cycles, so it raised the coefficient back. Write, meanwhile, surged to three new articles in one period by harvesting children from a structural gap analysis. See the Meta Report.

Metrics

• Total articles: 202
• Articles changed since last deploy: 10 (4 new + 2 edits + 1 structural rename + 1 meta + 1 research + 1 groom)

2026-04-10 (night)

What’s New

• New article: Research, Plan, Implement – a three-phase workflow that separates understanding from planning from execution, catching agent misunderstandings before they get cemented into code.
• New article: Platform as a Product – how to run your internal developer platform with the same discipline you’d use for a product with paying customers.
• New article: Strangler Fig – how to replace a legacy system incrementally by building new functionality alongside it, routing traffic piece by piece, until the old system can be switched off.
• Improved: The Aggregate article gained tighter prose and new cross-references to Invariant and Instruction File.
• Improved: The Architecture article gained tighter prose and new cross-references to Design Doc, Domain Model, Cognitive Load, and Coupling.
• Improved: The Feedback Loop article gained tighter prose and more varied sentence structures.
• Improved: The Determinism article gained tighter prose, prerequisite links, and a new cross-reference to Test.
• Improved: The Research, Plan, Implement article gained tighter prose and cleaner sentence structures.
• Improved: The Test-Driven Development article gained a new research finding on how agents use TDD, tighter prose, and cross-references to Requirement and Verification Loop.
• Improved: The Platform as a Product article gained tighter prose and better paragraph structure in the examples.
• Improved: The Hook article gained tighter prose, a new Sources section tracing the concept from the Gang of Four through Git and React to agentic harnesses, and a cross-reference to the Ralph Wiggum Loop failure mode.
• Sources: Added intellectual lineage to three articles: Plan Mode (ReAct through Plan-and-Execute to modern coding tools), Parallelization (Amdahl’s Law through Anthropic’s workflow taxonomy), and Determinism (Turing, Rabin and Scott, and Bernhardt’s “functional core, imperative shell”).
• Structural: Added 17 missing cross-reference backlinks to improve navigation between Design Heuristics articles and their related patterns across the encyclopedia.
• Meta: The engine detected a corpus stabilization signal as edit improvements shrink, reduced the sources coefficient as tracked coverage nears completion, and is watching the article pipeline as research goes quiet. See the Meta Report.

Metrics

• Total articles: 198
• Articles changed since last deploy: 22 (3 new + 8 edits + 3 sources audits + 1 groom + 2 meta + 1 research + sweep no-op)

2026-04-10 (evening)

What’s New

• New article: Feedback Loop – the architectural primitive that lets systems self-correct, from CI pipelines to agentic verification workflows.
• New article: Aggregate – how to cluster entities and value objects into consistency boundaries with a single root guarding all access.
• New article: Generator-Evaluator – split code creation and code critique into separate agents so neither role can blind the other.
• Improved: The Plan Mode article gained a concrete agent interaction transcript showing plan-then-execute in action.
• Improved: The Parallelization article gained a concrete interaction example showing three agents building independent API endpoints in parallel worktrees, then merging cleanly.
• Improved: The Context Engineering article gained a concrete interaction example showing how deliberate file selection and context ordering produce better agent output on the first try.
• Improved: The Verification Loop article gained a concrete interaction example showing an agent iterating through test failures after adding rate limiting.
• Improved: The Subagent article gained a concrete interaction example showing how a parent agent dispatches an exploration subagent and uses its compact summary to plan next steps.
• Improved: The A2A (Agent-to-Agent Protocol) article now covers the version 1.0 release, five production-ready SDKs, and tighter prose throughout.
• Improved: The Generator-Evaluator article gained tighter prose and the AgentCoder paper as a source for multi-agent code generation benchmarks.
• Improved: The Value Object article gained tighter prose and better paragraph flow.
• Improved: The Protocol article gained tighter prose and new cross-references to MCP and A2A.
• Sources: Added intellectual lineage to three articles: API (Joshua Bloch’s design principles and Roy Fielding’s REST dissertation), Algorithmic Complexity (Big-O notation from its 1894 origins through Knuth and Hartmanis-Stearns), and Protocol (TCP and the end-to-end principle through HTTP to MCP and A2A).
• Meta: The engine confirmed its edit-dominance prediction. A single critique pass added concrete interaction transcripts to four agentic articles in one period. See the Meta Report.

Metrics

• Total articles: 195
• Articles changed since last deploy: 20 (3 new + 9 edits + 3 sources audits + 1 groom + 2 meta + 1 research + 1 critique)

2026-04-10

What’s New

• New article: Value Object – the complement to Entity, covering objects defined by what they contain rather than who they are, with practical guidance for domain modeling and agentic workflows.
• New article: A2A (Agent-to-Agent Protocol) – how agents from different vendors discover each other’s capabilities and collaborate on tasks through a standard protocol.
• Improved: The Harness (Agentic) article gained tighter prose and Sources tracing the concept from Boeckeler’s “harness engineering” discipline to Russell & Norvig’s agent loop.
• Improved: The Retrieval article gained tighter prose, varied scenario framing, and clearer consequences.
• Improved: The Feedforward article gained clearer prose, added security scanning as a feedforward example, and fixed a source attribution.
• Improved: The Entity article gained tighter prose, better paragraph structure, and clearer guidance on telling agents which concepts carry identity.
• Sources: Added intellectual lineage to eight articles: Algorithm (al-Khwarizmi through Turing to Knuth), DRY (Hunt & Thomas through Beck to Codd), Product-Market Fit (Rachleff through Andreessen to Ellis and Blank), Prompt (GPT-3 few-shot through chain-of-thought), Separation of Concerns (Dijkstra through Parnas to Reenskaug’s MVC), Test (Dijkstra through Myers to Cohn and Beck), Tool (ReAct through Toolformer to OpenAI function calling), and Verification Loop (Wiener’s cybernetics through Beck’s TDD).
• Meta: The engine caught sources audits consuming half of all cycles and rebalanced, putting new article writing back in the lead. Separately, the action distribution reached its healthiest balance yet after a self-correcting edit plateau. See the Meta Report.

Metrics

• Total articles: 195
• Articles changed since last deploy: 20 (2 new + 4 edits + 8 sources audits + 1 groom + 2 meta + 1 research + 1 critique)

2026-04-09

What’s New

• New article: Retrieval – how agents pull relevant documents from an external corpus at query time, the pattern behind RAG.
• Improved: The MCP article now has corrected specification timeline (March and June 2025 releases separated) and coverage of the November 2025 spec release introducing the Tasks primitive.
• Improved: The Specification article gained a new section showing how to use thin written specs as runnable prototypes, with a worked example of a customer-deduplication tool built through three iterations.
• Improved: The Premature Optimization article gained a top-of-page “Understand This First” section linking to Performance Envelope and Observability.
• Improved: The Stream-Aligned Team article gained a clearer comparison between stream-aligned and component-aligned agent setups, with the payments-team scenario rewritten as a clean before/after.
• Improved: The Inverse Conway Maneuver article gained clearer paragraphs in the Solution section and a sharpened Consequences section that leads with the specific risk that agents won’t tell you when boundaries are wrong.
• Improved: The Coding Convention article gained paragraph splits in Problem, Solution, and How It Plays Out for easier reading.
• Sources: The Context Window article now traces the concept to the Transformer architecture paper, the “Lost in the Middle” research on positional attention bias, and the coining of “context engineering.”
• Structural: Added a full table of contents to the cover page to fix mobile navigation – visitors were landing on the cover, seeing only art and a paragraph, and bouncing because the sidebar TOC was hidden behind the hamburger menu. Also added a site-wide meta description for better search engine results.
• Meta: The engine spent 10 cycles editing without writing a single new article – its first write-free period – because clearing a draft backlog was mathematically more urgent. With only 2 drafts left (an all-time low), the balance shifted back toward new content. See the Meta Report.

Metrics

• Total articles: 193
• Articles changed since last deploy: 19 (1 new + 5 edits + 1 sources audit + 1 groom + 1 meta + cover TOC restructure)

2026-04-07

What’s New

• New article: Coding Convention – how to capture your team’s code style as a written, living artifact that both humans and AI agents can read and follow, and why the 2026 “naming renaissance” made this newly important.
• New article: Entity – how to recognize the things in your domain that have identity (customers, orders, invoices) and why keeping that identity stable matters when an AI agent is updating your code.
• New article: Stream-Aligned Team – why organizing teams around value streams instead of technical layers produces better software, and how the same principle applies when scoping AI agents.
• New article: Enabling Team – how temporary, teaching-oriented teams help stream-aligned teams adopt new capabilities like AI agent workflows without creating permanent dependencies.
• New article: Inverse Conway Maneuver – how to reshape your teams (or agents) to produce the software architecture you actually want, instead of the one your org chart would impose.
• Improved: The Specification article now shows how to use thin written specs as runnable prototypes that agents execute to surface unknown requirements, with a worked example following a team that discovers what their customer-deduplication tool actually needs to do.
• Improved: The Memory article gained decay heuristics that keep memories self-maintaining, automated extraction that harvests lessons from your conversation history, and a cold-start guide for the first week of use.
• Improved: The Ralph Wiggum Loop article gained five named failure modes and their fixes, so you can diagnose what went wrong when your loop stops converging.
• Improved: The Compaction article gained new detail on how harnesses configure automatic compaction (reserve token thresholds, API endpoints) and a clearer explanation of the tradeoff between automatic and manual triggers.
• Improved: The Bounded Autonomy article gained a longer-horizon scenario showing how early restrictive policies evolve into earned autonomy, plus new guidance on treating tier regressions as a healthy feedback signal rather than a setback.
• Improved: The Approval Fatigue article gained a sharper opening, a connection to the forty-year-old “alert fatigue” pattern from security operations, and a new cross-reference to Blast Radius.
• Improved: The Shadow Agent article gained a sharper intent line and a clearer explanation of what makes shadow agents worse than shadow IT (they take actions, not just hold data).
• Improved: The Big Ball of Mud article gained a sharper intent line, smoother prose, and Martin Fowler’s Strangler Fig pattern in the Sources.
• Improved: The Metric article gained tighter prose, sharper analysis of why traditional metrics mislead in AI-assisted workflows, and cross-references to Verification Loop and Instruction File.
• Sources: The Architecture article now credits Perry and Wolf for founding the field, Shaw and Garlan for the first textbook, Martin Fowler and Ralph Johnson for the modern “decisions costly to reverse” framing, and Christopher Alexander for the pattern-language approach. The Continuous Integration article credits Grady Booch for coining the term, Kent Beck for formalizing the practice in Extreme Programming, and Jez Humble and David Farley for extending CI into continuous delivery. The Eval article credits OpenAI’s Evals framework, the HumanEval benchmark, and Princeton’s SWE-bench as foundational contributions to how we measure agent performance.
• Structural: The Approval Fatigue, Shadow Agent, Technical Debt, and Big Ball of Mud antipatterns now appear on their section landing pages, where readers browsing a topic area can discover them alongside the patterns they complement.
• Meta: The engine found a bookkeeping bug where 41 articles had Sources sections in their files but weren’t marked as audited in state, causing the audit action to re-pick articles whose work was already done – now fixed with a backfill and new rules that prevent future drift. See the Meta Report.

Metrics

• Total articles: 192
• Coverage: 192 of 217 proposed concepts written (88%)
• Articles changed since last deploy: 18 (5 new + 9 edits + 3 sources audits + 1 groom + 1 meta backfill)

2026-04-06 (afternoon)

What’s New

• New article: Architecture Fitness Function – automated checks that catch structural drift before it compounds, keeping your architecture aligned with intent.
• New article: Ownership – who is responsible for this code, and what happens when nobody can answer that question.
• New article: RAG Poisoning – how attackers corrupt the knowledge bases AI agents retrieve from, causing agents to treat fabricated information as verified fact.
• New article: Shift-Left Feedback – move quality checks earlier in the agent’s workflow so mistakes are caught while they’re still cheap to fix.
• New article: Metric – what makes a number worth tracking, why AI-generated code makes measurement more important than ever, and how to avoid vanity metrics.
• Improved: The Garbage Collection article gained empirical data on agent-caused code drift, a new measurement component, and a third scenario showing how sweep logs surface root causes.
• Improved: The Agent Trap article added a section on dynamic cloaking attacks and notes on the unresolved legal liability question.
• Improved: The Vibe Coding article added the concept of comprehension debt, new CVE data, and a third scenario showing how vibe-coded projects become unmaintainable.
• Improved: The Model Routing article added cascading as a routing strategy and restructured the examples for clarity.
• Improved: The Printf Debugging article added a stronger explanation of why this is the default debugging method for AI coding agents.
• Improved: The Team Cognitive Load, Architecture Fitness Function, RAG Poisoning, Ownership, and Shift-Left Feedback articles received prose quality passes.
• Sources: The YAGNI article now traces the principle back to Kent Beck’s C3 project and the Extreme Programming community. The Prompt Injection article credits the researchers who discovered and named the vulnerability.
• Meta: The engine’s write throughput hit an all-time high after rebalancing, but sources audits dropped to zero – adjusted the weight to find the sweet spot. See the Meta Report.

Metrics

• Total articles: 187
• Coverage: 187 of 216 proposed concepts written (87%)
• Articles changed since last deploy: 19 (5 new + 10 edits + 2 sources audits + 2 meta)

2026-04-06 (overnight)

What’s New

• New article: Garbage Collection – how recurring agent-driven sweeps keep a codebase from drifting away from its own standards.
• New article: Agent Trap – the umbrella concept for adversarial content that exploits AI agents by corrupting their environment rather than attacking the model itself.
• New article: Vibe Coding – the most talked-about antipattern in agentic coding, and why generating code you don’t understand is a trap that works until it doesn’t.
• New article: Model Routing – how to match AI models to tasks so you spend your budget where it matters.
• New article: Printf Debugging – the oldest and most universal debugging technique, and the one AI coding agents reach for instinctively.
• New article: Best Current Practice – every recommendation carries its own expiration warning, and this matters when AI agents trained on older data may suggest stale approaches.
• Improved: The Agent Teams article gained a new Orchestration Topologies section covering the four coordination patterns used in production multi-agent systems.
• Improved: The Logging article gained a section on logging at boundaries – the highest-value instrumentation points in any system.
• Improved: The Happy Path article added alternative names (Golden Path, Sunny Day Scenario) and better paragraph structure.
• Improved: The Ralph Wiggum Loop, Bounded Context, and Externalized State articles received prose quality passes.
• Sources: The Red/Green TDD, Cohesion, and Subagent articles now credit their intellectual origins.
• Meta: The engine’s action mix rebalanced after a coefficient experiment, and a dormant action type is being retired after 40+ idle cycles. See the Meta Report.

Metrics

• Total articles: 171
• Coverage: 171 of 209 proposed concepts written (82%)
• Articles changed since last deploy: 15 (6 new + 6 edits + 3 sources audits)

2026-04-06 (late night)

What’s New

• New article: Approval Fatigue – when approval requests come faster than a human can genuinely evaluate them, oversight degrades into rubber-stamping.
• New article: Shadow Agent – an AI agent operating inside your organization without anyone in governance knowing it exists.
• New article: Tool Poisoning – malicious tool descriptions that hijack agent behavior through the tool discovery channel.
• New article: Big Ball of Mud – the most common software architecture in practice: a haphazardly structured, sprawling, duct-taped system that resists all attempts at understanding.
• New article: Premature Optimization – spending effort to make code faster before you know where the bottleneck is, trading clarity for speed that nobody needed.
• New article: Technical Debt – the accumulated cost of shortcuts, deferred maintenance, and expedient decisions that make future changes harder and slower.
• Visual: Antipattern articles now display a distinctive red prohibition sign admonishment, distinguishing them from patterns (green checkmark) and concepts (blue info icon).
• Improved: Local graph widget now shows inverted labels for incoming edges, making relationship direction clearer.
• Cross-references: Added reciprocal Related Patterns links across 30 existing articles connecting them to the six new antipatterns.

Metrics

• Total articles: 165
• Coverage: 165 of 206 proposed concepts written (80%)
• Articles edited since last deploy: 37 (6 new antipatterns + 30 reciprocal link updates + 1 graph fix)

2026-04-06 (evening)

What’s New

• New feature: Pattern Map – an interactive force-directed knowledge graph showing all 159 articles and 893 connections. Search, zoom, hover to highlight connections, drag to rearrange, click to navigate. Every article page now has a local graph widget below the type marker showing its immediate neighbors.
• New article: Team Cognitive Load – how the mental effort of understanding and maintaining systems limits what teams and agents can effectively own.
• New article: Ralph Wiggum Loop – the embarrassingly simple pattern of restarting an agent with fresh context after each unit of work, using a plan file instead of an orchestration framework.
• New article: Happy Path – the default scenario where everything works, and why recognizing it is the first step toward building software that handles the real world.
• Improved: The Checkpoint article gained coverage of ephemeral environments as a checkpoint strategy. The Bounded Autonomy article now covers dynamic trust-score de-escalation. The Model article now covers reasoning capabilities, multimodal input, and model selection guidance. The MCP article now covers Linux Foundation governance, Streamable HTTP transport, and OAuth 2.1 authentication.
• Sources: Added intellectual lineage to the Code Smell, Agent, and AI Smell articles.
• Infrastructure: Social preview images for link sharing, external links open in new tabs, site now indexable by search engines, sitemap live for crawlers.
• Other: Updated the Meta Report with the engine’s eighth self-evaluation.

Metrics

• Total articles: 159
• Coverage: 159 of 191 proposed concepts written (83%)
• Articles edited since last deploy: 15 (3 new articles + 4 targeted edits + 3 sources audits + 5 infrastructure)

2026-04-06 (morning)

What’s New

• New article: Ralph Wiggum Loop – the embarrassingly simple pattern of restarting an agent with fresh context after each unit of work, using a plan file instead of an orchestration framework.
• New article: Agent Teams – how multiple AI agents coordinate through shared task lists and peer messaging, scaling agentic work beyond what one human can direct.
• New article: Externalized State – how to store an agent’s plan, progress, and intermediate results in files so workflows survive interruptions and stay auditable.
• New article: Logging – how to record what your software does as it runs, covering structured logs, severity levels, and why logging is the primary way both humans and AI agents understand runtime behavior.
• New article: Happy Path – the default scenario where everything works, and why recognizing it is the first step toward building software that handles the real world.
• Improved: The Context Engineering article now covers four named operations (select, compress, order, isolate), signal-to-noise framing, and production-scale concerns like cache efficiency.
• Improved: The MCP article now covers current governance (Linux Foundation), Streamable HTTP transport, OAuth 2.1 authentication, security threats, and adoption metrics.
• Improved: The Model article now covers reasoning capabilities, multimodal input, model selection guidance, and intellectual sources.
• Improved: The Subagent article gained three named use case categories (exploration, parallel processing, specialist roles), a warning against overuse, and guidance on using cheaper models for subagent tasks.
• Improved: The Agent article gained cross-section links to Least Privilege, Boundary, and Test, connecting the book’s central agentic concept to foundational patterns.
• Improved: The AI Smell article gained a new section on agent struggle as a code quality signal – when your agent fails repeatedly, the problem may be your code, not the agent.
• Improved: The Steering Loop article gained tighter prose, a new section on completion gates, and proper source attribution.
• Improved: The Bounded Autonomy article gained tighter prose and added coverage of dynamic trust-score de-escalation.
• Improved: The Checkpoint, Design Doc, Architecture Decision Record, and Conway’s Law articles received prose quality improvements.
• Improved: Added intellectual lineage to the Crossing the Chasm and Skill articles.
• Other: Updated the Meta Report with the engine’s seventh self-evaluation: both previous hypotheses confirmed, coverage velocity doubled, and the new stochastic selection system shows early promise.

Metrics

• Total articles: 165
• Coverage: 165 of 200 proposed concepts written (83%)
• Articles edited since last deploy: 19 (5 new articles + 12 targeted edits + 2 sources audits)

2026-04-06

What’s New

• New article: Checkpoint – how to insert verification gates into agentic workflows so agents catch errors at each stage instead of building on broken foundations.
• New article: Architecture Decision Record – how to capture design decisions so future readers (human or AI) don’t have to guess why the system is built this way.
• Improved: Every article now displays a visual marker identifying it as either a Pattern (a solution you can apply) or a Concept (an idea to recognize and understand), helping readers orient instantly.
• Improved: The Feedback Sensor article received tighter prose, a new Sources section, and stronger motivation for why automated checks matter.
• Improved: Added a Sources section to the Memory article, tracing the concept’s origins from cognitive psychology through modern AI agent memory systems.
• Other: Updated the Meta Report with the engine’s sixth self-evaluation: all signals stable or improving, no process changes needed.

Metrics

• Total articles: 153
• Coverage: 153 of 188 proposed concepts written (81%)
• Articles edited since last deploy: 156 (2 new articles + 1 targeted edit + 1 sources audit + 152 via entry type markers sweep)

2026-04-05 (late)

What’s New

• New article: Bounded Autonomy – how to calibrate agent freedom based on the consequence and reversibility of each action, from full autonomy for safe tasks to human-only for critical operations.
• Improved: The Naming article received tighter prose, proper source attribution crediting Robert C. Martin and Phil Karlton, and a clearer presentation of naming principles.
• Improved: The Refactor article now credits the people who originated the ideas it teaches – from Opdyke and Johnson coining the term in 1992, through Fowler’s canonical catalog, to Beck’s integration with testing.
• Structural: Section index pages for Socio-Technical Systems and Agent Governance and Feedback now show a “Work in Progress” notice indicating more entries are on the way.
• Other: Updated the Meta Report with the engine’s fifth self-evaluation: the draft-pressure fix is confirmed working, and the restructure action’s weight continues its planned decay.

Metrics

• Total articles: 158
• Coverage: 158 of 192 proposed concepts written (82%)
• Articles edited since last deploy: 3 (1 targeted edit + 1 sources audit + 1 new article)

2026-04-06

What’s New

• New article: Design Doc – how to translate requirements into a technical plan before building starts, and why this matters even more when an AI agent is the builder.
• Improved: The Skill article gained a new section on how skills evolve from ad-hoc instructions into reliable team workflows, plus a new scenario showing code review skill evolution in practice.
• Improved: The Ubiquitous Language article received proper source attribution, tighter prose in the agentic workflow section, and a new cross-link to the Instruction File pattern.
• Improved: Added intellectual lineage to the Feedforward article, tracing the concept from 1920s control theory through Marshall Goldsmith’s coaching framework to Birgitta Boeckeler’s guides-and-sensors model.
• Structural: Improved cross-reference navigation in the Security and Trust section – 14 missing reciprocal links added so readers can follow connections in both directions.
• Other: Updated the Meta Report with the engine’s fourth self-evaluation: a procedural bug was keeping unreviewed articles from getting edited, now fixed with a clearer priority gate.

Metrics

• Total articles: 158
• Coverage: 158 of 189 proposed concepts written (84%)
• Articles edited since last deploy: 10 (2 targeted edits + 1 sources audit + 1 groom pass across 6 articles)

2026-04-05 (evening)

What’s New

• New article: Conway’s Law – why software systems end up mirroring the communication structure of the teams that build them, and how to use this force deliberately when organizing both human teams and AI agents. This is the first article in the new Socio-Technical Systems section.
• Improved: Updated the Prompt Injection article with 2025-2026 developments: direct vs. indirect injection, MCP attack surfaces, instruction hierarchy defenses, multimodal vectors, and detection techniques like canary tokens.
• Improved: Every pattern entry now shows prerequisite concepts at the top of the page – follow the links to drill down to foundational ideas before reading advanced ones.
• Improved: The Test-Driven Development article now credits Kent Beck, the Extreme Programming community, Robert C. Martin, and Martin Fowler for the ideas it teaches.
• Structural: Fixed paragraph line spacing to match the intended readability standard across all article pages.
• Other: Updated the Meta Report with the engine’s second self-evaluation: the rotation rebalancing worked, all three hypotheses were resolved, and a course correction prevents the idea pipeline from drying up.

Metrics

• Total articles: 149
• Coverage: 149 of 169 proposed concepts written (88%)
• Articles edited since last deploy: 107 (2 targeted edits + 1 sources audit + 104 via Understand This First sweep)

2026-04-05

What’s New

• New article: Domain Model – how to capture the concepts, rules, and relationships of a business problem so that both humans and AI agents share the same understanding.
• New article: Ubiquitous Language – how a shared vocabulary drawn from the business domain keeps developers, stakeholders, and AI agents aligned on what every term means.
• New article: Naming – how choosing clear, consistent identifiers for code elements matters more in the agent era, where AI amplifies whatever naming patterns it finds.
• New article: Bounded Context – how drawing explicit boundaries around parts of your system keeps domain models focused and prevents vocabulary collisions, especially when directing AI agents.
• New article: Feedforward – how to steer an AI agent toward correct output before it acts, using instruction files, specifications, and computational checks.
• New article: Feedback Sensor – how automated checks after each agent action detect mistakes and drive self-correction, from fast type checkers to LLM-based code reviewers.
• New article: Steering Loop – how the closed cycle of act, sense, and adjust turns feedforward controls and feedback sensors into a system that converges on correct code.
• New article: Harnessability – why some codebases are easier for AI agents to work in than others, and how type systems, module boundaries, and codified conventions determine the ceiling on agent effectiveness.
• Improved: Added example prompts to 129 pattern entries, showing readers what it looks like to apply each concept when directing an AI coding agent.
• Improved: The Harnessability article gained a practical optimization checklist – six concrete steps to make your codebase more tractable for AI agents.
• Improved: The Domain Model article gained a new section on encoding behavior in domain objects, tighter prose, a corrected alias, and a Sources section crediting Eric Evans and Martin Fowler.
• Improved: The Feedforward article received tighter prose and a corrected reference link.
• Other: Published the first Meta Report entry, documenting how the improvement engine measures and adjusts its own process.

Metrics

• Total articles: 155
• Coverage: 155 of 178 proposed concepts written (87%)
• Articles edited since last deploy: 132 (4 targeted edits + 1 sources audit + 129 via example-prompts sweep)

2026-04-04

What’s New

• New article: Specification covers how to write what a system should do precisely enough for a human or an agent to build it correctly.
• Improved: The Specification article received tighter prose, a unique epigraph, and new content on the three levels of spec-driven development.
• Improved: Five core agentic coding articles (Context Window, Context Engineering, Prompt, Agent, Tool) now include example prompts showing how to apply each pattern when directing an AI agent.

Metrics

• Total articles: 140
• Coverage: 140 of 200 proposed concepts written (70%)
• Articles edited since last deploy: 6