AGENTS.md: Why Humans Should Write Agent Context (Not AI)
A study from ETH Zurich shows that agents writing their own context files reduce performance by 3% and increase costs by 20%. See the tested Autenticare template.
Fabiano Brito
CTO, Autenticare
AGENTS.md to the LLM itself seems efficient, but costs dearly in predictability. Recent data from ETH Zurich (Gloaguen et al.) proves: a context file rigorously maintained by humans outperforms self-generation in performance and token economy.
Orchestrating autonomous agents in complex repositories requires clear rules. When you let agents loose in the code, they need a geographical map of the project, architectural boundaries, and local anti-patterns. The most common solution for this in the community is the AGENTS.md file.
But who should write this file?
Many teams automate context creation via LLMs (“read this repo and summarize the rules”). However, research and our practical experience at Autenticare show that this is an architectural flaw.
The Cost of Self-Generation
resolution success rate
increase in token consumption
tested ideal model
According to the recent study from ETH Zurich, allowing agents to infer and document project rules by themselves leads to a 3% decrease in success of task execution and a 20% increase in costs.
The reason is simple: the LLM focuses on “what” the code does (which is already obvious by reading the files), but lacks the intentionality of “why” it was done that way. An agent generating AGENTS.md creates empty technical descriptions instead of business constraints.
The Tested Autenticare Template
Define the expected behavior, tone of voice (such as the mandatory use of direct responses), and the boundaries of what the agent can or cannot do in the repo.
Instead of describing what the src/ folder is, establish WHERE new things should go (e.g.: "Never create generic components outside of ui/").
The most valuable part: list the solutions that seem correct but have been banned in your company (e.g.: "Never use lodash, we have native helpers").
AGENTS.md file is your contract with the AI. If the AI writes its own contract, it naturally shapes the rules towards paths of least resistance, ignoring long-term technical debt.
For scalable agentic engineering, write the strategic context by hand and let the LLM automate only the tactical execution.
Your AGENTS.md is not a wiki; it is a behavioral electric fence. Treat it as such.
To structure the foundation of your AI orchestration with Autenticare’s best practices, discover our Agent Factory. Need help with your platform’s governance? Get in contact with our team.