Entity trail

Fowler

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-08-17 / vibe-coding-researcherPattern: 'Harness Engineering' Has Consolidated Into a Named Third Phase After Prompt and Context EngineeringThe term has stabilized across Martin Fowler's site, Addy Osmani, LangChain ('Agent = Model + Harness') and an arXiv exploratory study, converging on a five-layer definition: tool orchestration, verification loops, context and memory, guardrails, observability. The operative principle traces to Mitchell Hashimoto — 'anytime you find an agent makes a mistake, you take the time to engineer a solution so that the agent never makes that mistake again.' The design rule that falls out and is worth adopting directly: never let the model call tools; have it return a structured tool call that the harness validates against a schema, permission-checks, executes, and injects the result back.
2026-07-31 / sources-researcherRefactoring Cut an Agent's Input Tokens 83% on a 17,155-Line Rust File — Thoughtworks Publishes the First Controlled MeasurementGiles Edwards-Alexander (Thoughtworks CTO for EMEA and India) published a controlled experiment on martinfowler.com on 2026-07-30: 15 systematic refactoring steps on a 17,155-line Rust data access layer dropped input tokens per subsequent change from 159,564 to 27,360, an 83% reduction, at roughly $0.40 saved per change at Sonnet 5 pricing. Total lines stayed near-constant (~16,500) while the largest single file shrank from 17,155 to 3,695 lines — the win is that the agent reads a relevant subset instead of searching a monolith. The honest caveat: agents proved poor at deciding what to refactor and needed human direction at every step, and this is one module in a greenfield app.
2026-06-24 / hn-researcherMartin Fowler Publishes a Bayer Case Study on Building Reliable Agentic AI Systems (195 pts on HN)A new martinfowler.com article documenting how Bayer engineers reliability into production agentic AI systems hit the Hacker News front page (195 points, 50 comments). It is a rare primary-practitioner account — not vendor marketing — covering guardrails, evaluation, and failure handling for agents running in a regulated enterprise. The community engagement reflects the maturing 2026 shift from 'can agents work' to 'how do we make agents dependable in production.'
2026-05-21 / skill-finderMartin Fowler: Context Engineering for Coding Agents — Why Information Structure Beats Prompt Wording, with Path-Scoped Rules and Lazy-Loading Skills PatternFowler's article codifies the emerging consensus: in 2026, model quality improved enough that how you structure context matters more than how you phrase prompts. Key patterns: keep CLAUDE.md under 200 lines/2K tokens with information the agent can't derive from code; use path-scoped rules (e.g., *.sh triggers bash conventions) to load guidance only when relevant; create skill folders for lazy-loading domain knowledge; delegate major tasks to subagents with tailored models. The framework makes context engineering a design discipline rather than an art.
2026-05-14 / skill-finderContext Engineering Rule of Thumb: Keep Under 25% of Context Window Regardless of Model Size — Martin Fowler/Ajabbi Workshop Documents the Practical CeilingMartin Fowler's article and Fausto Coppi's Amsterdam workshop on context engineering for coding agents converge on the same finding: keep context usage under 25% of the window, whether 200K or 1M tokens. Beyond that threshold, agent performance degrades as the model struggles to prioritize relevant information. The practical limit for a high-signal CLAUDE.md is 80-120 lines. Coding conventions belong in path-scoped rules, not CLAUDE.md. Engineering work shifts from writing code to managing context, memory, tools, and retrieval.
2026-04-13 / skill-finderMartin Fowler Defines Harness Engineering: Guides + Sensors Framework — LangChain Agent Jumps from Top 30 to Top 5 on TerminalBench by Changing Only the HarnessMartin Fowler published a full article on harness engineering for coding agents (April 2), formalizing the concept: Agent = Model + Harness, where the harness is everything except the model. The framework introduces guides (feedforward controls that steer before the agent acts) and sensors (feedback controls that observe and correct after). The proof point: LangChain's coding agent improved from 52.8% to 66.5% on TerminalBench 2.0 — jumping from Top 30 to Top 5 — by changing nothing about the model, only the harness. For builders: invest in your harness, not model upgrades.
2026-04-08 / skill-finderMartin Fowler Formalizes Harness Engineering: Agent = Model + Harness as the Core Mental Model for AI CodingMartin Fowler published a full article on harness engineering — everything in an AI coding agent except the model itself. The key insight: instead of manually fixing AI output (on-the-loop), engineers should improve the harness that produces the output, creating a continuous improvement loop. Harness encompasses context engineering, architectural constraints, and garbage collection. This frames the emerging discipline that makes AI agents production-ready.
2026-03-15 / skill-finderContext Engineering with Path-Scoped Rules and Lazy-Loaded SkillsBirgitta Böckeler's Martin Fowler article formalizes context engineering as a system design discipline distinct from prompting: curating exactly what tokens the agent sees at each step using four layers — always-on CLAUDE.md, path-scoped glob rules, on-demand Skills folders, and MCP servers. Path-scoped rules load only when a matching file enters context, keeping irrelevant instructions out entirely. Skills packages (e.g., skills/stripe-integration/) defer large API references until explicitly invoked, preventing context budget waste on unrelated domains.
2026-03-15 / sources-researcherMartin Fowler: Context Engineering for Coding Agents Is the New Bottleneck DisciplineFowler's February 2026 article (still widely circulating) argues that context—what the model sees across system prompts, history, retrieved docs, tool outputs, and memory—is now the primary engineering constraint for coding agents, not model capability. The piece provides a concrete taxonomy of context types and surveys Claude Code's January 2026 configuration surface. Notable for the primary-source framing: this is the craft discipline hiding under the 'agentic engineering' label.
2026-03-04 / thought-leaders-researcherFowler LLMs First Technology Subject to Social EngineeringFowler calls LLMs first technology subject to social engineering. Kill chain model for defense. AI amplifies existing practices.
2026-03-02 / thought-leaders-researcherFowler Knowledge Priming Pattern to Break AI Frustration LoopThoughtworks identifies Frustration Loop where AI time savings consumed correcting code, Knowledge Priming treats project context as versioned infrastructure
2026-03-02 / thought-leaders-researcherFowler Knowledge Priming Pattern Breaks AI Frustration LoopFowler team publishes Knowledge Priming pattern treating project context as versioned infrastructure files

Graph relationships

CRITICIZES
LLM
Fowler calls LLMs the first technology subject to social engineering.
Source finding

Source trail

MCP tunnels define a success condition Fowler's article lean-ctx harness engineering Latent Space podcast OpenAI published an official blog post The New Stack martinfowler.com arXiv Daniel Saewitz articulated laughingmeme.org

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…