MindPattern
← The Wire

Top 5 · 2026-04-13 · source-backed

Martin Fowler Just Gave Harness Engineering a Name, and the Numbers Back Him Up

Confidence
source-backed
Sources
1
Redaction
passed

Story

Martin Fowler published a full article on April 2 formalizing something I've been feeling for months: the thing that separates a good coding agent from a bad one isn't the model. It's everything around the model. He calls it harness engineering.

The framework is clean. Agent = Model + Harness. The harness is every piece of scaffolding you build around the LLM: the system prompts, the file selection logic, the retry strategy, the output validators, the context window management. Fowler breaks it into two control types. Guides are feedforward controls that steer the agent before it acts (think CLAUDE.md files, system prompts, structured output schemas). Sensors are feedback controls that observe and correct after the agent acts (think linters, test suites, diff reviews, approval gates).

The proof point is what makes this more than theory. LangChain's coding agent improved from 52.8% to 66.5% on TerminalBench 2.0, jumping from somewhere in the Top 30 to Top 5. They changed zero model code. Zero. Every point of improvement came from harness changes: better context selection, smarter retry logic, improved output parsing.

This matches my experience exactly. I've spent the last year building MindPattern, and I'd estimate 80% of my engineering time goes into harness work, not prompt engineering. The CLAUDE.md files, the agent dispatch logic, the quality gates, the synthesis passes. The model is the engine, but the harness is the car.

For builders, this gives us a shared vocabulary that's been missing. When someone asks "how do I make my coding agent better?" the answer is almost never "use a better model." It's "improve your harness." Specifically: audit your guides (are you giving the agent enough context before it acts?) and your sensors (are you catching failures quickly enough to correct course?). If you're running Claude Code, your CLAUDE.md file IS your primary guide. Your test suite IS your primary sensor. Invest there.

I think Fowler naming this is going to accelerate the field. The same way "DevOps" gave infrastructure engineers a professional identity, "harness engineering" gives agent builders one. We're not prompt engineers. We're harness engineers.

Related stories

Each link below shares sources, entities, or timing with this story.

  1. Shared entities / Same source / Shared topic / Earlier coverage

    Three Engineers. One Million Lines of Code. Zero Human-Written. Harness Engineering Is Now a Discipline.

    Both cover Agent, Claude Code, Fowler, Harness; cite the same source (harness engineering); overlapping topics (agent, engineer, engineering, harness, output).

  2. Shared entities / Shared topic / Earlier coverage / Tension

    Sebastian Raschka Says Claude Code's Real Secret Sauce Is the Harness, Not the Model. The Leaked Source Proves It.

    Both cover Claude Code, Harness, LLM, Model; overlapping topics (agent, claude, coding, context, engineering); earlier Claude Code coverage from 2026-04-01.

  3. Shared entities / Shared topic / What happened next

    Anthropic studied 400,000 Claude Code sessions and found the thing you bring isn't code

    Both cover April, Claude, Claude Code, When; overlapping topics (better, claude, coding, engineer, model); picks up the April thread on 2026-06-23.

  4. Shared entities / Shared topic / What happened next / Tension

    Uber Burned Its Entire 2026 Claude Code Budget by April. The COO Says He Can't Justify the Spend.

    Both cover April, Claude Code, Harness, When; overlapping topics (claude, coding, engineer, engineering); picks up the April thread on 2026-05-26.

  5. Shared entities / Shared topic / What happened next

    Datadog Just Showed Us What AI in Production Actually Looks Like

    Both cover Claude, Claude Code, LangChain, LLM; overlapping topics (claude, prompt); picks up the Claude thread on 2026-05-02.

  6. Shared entities / Shared topic / What happened next / Tension

    Cursor 3 Ditches VS Code for Agent Orchestration. Claude Code Holds 54% Market Share at $1.2B ARR.

    Both cover Agent, CLAUDE, Claude Code, When; overlapping topics (agent, claude, coding); picks up the Agent thread on 2026-04-16.

  7. Shared entities / Shared topic / What happened next

    Karpathy's CLAUDE.md Template Grows to Ten Rules

    Both cover Claude, Claude Code, LLM, Name; overlapping topics (agent, claude, coding, engineering); picks up the Claude thread on 2026-06-30.

  8. RTK Just Hit 30K Stars Because Everyone's Token Bill Is Out of Control

    Both cover April, Claude Code, LLM, Zero; overlapping topics (claude, coding, context, output); picks up the April thread on 2026-04-20.

Source trail

harness engineering

Entities

AgentAprilCLAUDEClaude CodeDevOpsFowlerGuidesHarnessInvestLangChainLLMMindPatternModelNameSensorsSpecifically

Provenance

Canonical issue
Ramsay Research Agent — April 13, 2026
AI generated
no
Story unit
2026-04-13-martin-fowler-just-gave-harness-engineering-a-name-and-the-numbers-back-him-up
Labels
source-backed, canonical briefing excerpt
Martin Fowler Just Gave Harness Engineering a Name, and the Numbers Back Him Up | MindPattern