One AGENTS.md for four AI tools is harness debt
A single context file behaves differently in every harness and every model. The four kinds of harness debt we keep finding in teams, and what a Dark Factory does instead.
Written by
-
Benedikt Stemmildt Founder & Co-CEO
Published: 2 April 2026 Updated: 2 min read
Your team has one AGENTS.md. Five engineers. Four different AI tools. Zero of them behaving the same way.
Welcome to harness debt.
One blueprint for every factory floor
Teams check a single context file into their repo and assume it works everywhere. Claude Code, Cursor, Copilot, Windsurf. Claude, GPT, Gemini. One blueprint for every factory floor.
It doesn’t work. Same weights, same training, different harness, different results.
The blueprint was never universal.
The four kinds of harness debt we keep finding in teams
Harness debt
Same file, different tools. Each harness has different context windows, injection points and tool systems. What works in Claude Code degrades in Cursor without any warning. The wrong instruction in the wrong harness means amplified damage at scale.
Model debt
Prompts tuned for one model confuse another. Claude follows long instruction chains precisely. GPT rewrites your logic. Gemini suggests three alternatives. Same input, different factory output. And every model update reshuffles the deck.
Drift debt
Context files describe a codebase from three months ago. Context files grow by small additions and rarely lose what went stale. The agent reads stale blueprints and produces confident garbage.
Coupling debt
Architecture decisions, coding standards, agent instructions and business rules all jammed into one file. In a study of 2,303 context files, CLAUDE.md files scored a median 16.6 on the Flesch reading ease scale. That’s dense legal contract territory. When anything changes, the file stays untouched, because the team can’t parse it.
Why harness debt multiplies faster than traditional debt
Traditional tech debt compounds linearly. Harness debt multiplies: tools x models x team members x file staleness.
What a Dark Factory does differently
- Harness-specific context. One source of truth per tool, not one file pretending to fit all.
- Model-aware instructions. What Claude needs is not what GPT needs. Engineer for each.
- Context files owned like code. Reviewed, versioned, tested against actual agent behavior.
- Decoupled layers. Architecture decisions, coding standards and agent instructions in separate files, so you change one without breaking the others.
- Quarterly harness audits, because whether your context drifted is a question of when, not if.
You can’t build a Dark Factory with a single AGENTS.md and hope for the best. That’s a Bullshit Factory with good intentions.
Are you engineering context per harness? Or copy-pasting one file and hoping it works?
(Source: Agent READMEs: An Empirical Study of Context Files for Agentic Coding (opens in a new tab), arXiv, November 2025.)
Further reading
Written by
-
Benedikt Stemmildt · Founder & Co-CEO
More on Context engineering
- Context engineering is the standardization you could never sell Preparing context for AI agents forces the documentation and standardization architecture teams have pushed for years. The agent does not need it more than the organization did. It removes the resistance.
- A study says context files hurt agents. It tested the wrong thing ETH Zurich found that auto-generated context files make agents worse. That shows auto-generation is useless, not context files. A useful context file records where the model fails your project.
- Context management matters more than picking the model Letting developers pick a model solves the wrong problem. The real constraint for coding agents is context, and AMP answers it with sub-agents, two working modes and fewer, sharper tools.
Test the constraint in your system
The Factory Readiness Assessment connects this research to your value stream, your data, and the next decision.
Training on this topic: Agentic Engineering for Developers and QAWant to check one team first? The self-check: twelve yes or no questions, three minutes, instant result, no account. Take the self-check