New ETH Zurich Study Proves Your AI Coding Agents Are Failing Because Your AGENTS.md Files Are Too Detailed

Why It Matters

The findings expose a hidden efficiency trap for enterprises deploying LLM‑powered coding assistants, directly impacting productivity and cloud spend. Optimizing context files can unlock faster, cheaper, and more reliable AI‑assisted development pipelines.

Key Takeaways

• Auto-generated context files cut success rates by ~3%
• Including context adds >20% inference cost and extra steps
• Human-written files improve performance only marginally, ~4% gain
• Detailed directory trees and style guides waste tokens without benefit
• Keep AGENTS.md under 60 lines; use pointers, not copies

Pulse Analysis

Context engineering has become a buzzword as teams strive to squeeze every ounce of performance from large language models. The promise of a single repository‑level file—AGENTS.md—was to give AI agents a reliable north star, consolidating architecture overviews, tooling choices, and coding conventions. In practice, however, the ETH Zurich research shows that this one‑size‑fits‑all approach often backfires. By injecting hundreds of extra tokens into each prompt, developers unintentionally create noise that competes with the model's internal knowledge, leading to slower reasoning and higher inference bills.

The study’s quantitative results highlight a stark trade‑off: auto‑generated context files shave roughly three percent off task success while inflating compute costs by more than twenty percent. Even carefully curated human files deliver only a modest four‑percent uplift. This paradox stems from the agents’ obedience to explicit instructions; when those directives are redundant—such as exhaustive directory trees or style guidelines—the model wastes valuable context windows on parsing irrelevant data. Consequently, the agents spend more reasoning steps navigating a fabricated manual rather than leveraging their built‑in code‑understanding capabilities.

Practitioners can turn these insights into actionable improvements. The research advocates a surgical, rather than comprehensive, mindset: include only the technical stack, intent, and non‑obvious tooling, and exclude verbose file listings or stylistic mandates. Keeping AGENTS.md under sixty lines, using pointers to live code, and employing progressive disclosure ensures each token adds measurable value. As LLMs continue to mature, disciplined context engineering will become a competitive differentiator, enabling faster development cycles, reduced cloud spend, and more trustworthy AI‑driven code generation.