- Updated: February 26, 2026
- 5 min read
ETH Zurich Study Reveals Overly Detailed AGENTS.md Files Undermine AI Coding Agents
The ETH Zurich study proves that overly detailed AGENTS.md files actually degrade AI coding agents’ success rates, raise inference costs by more than 20 % and add unnecessary reasoning steps.
ETH Zurich Study Reveals Why Your AI Coding Agents Fail: The Hidden Cost of Over‑Engineered AGENTS.md Files
What the ETH Zurich Researchers Discovered
A team of computer‑science scholars at ETH Zurich examined the performance of several state‑of‑the‑art coding agents—including Sonnet‑4.5, GPT‑5.2 and Qwen‑3‑30B—across the newly introduced AGENTBENCH benchmark. Their findings, published on arXiv, show a clear pattern: context files that are too verbose or automatically generated actually hurt the agents. Instead of acting as a “North Star,” these files become a source of noise, forcing the model to waste tokens on irrelevant instructions.
For developers, AI product managers and marketers who rely on AI agents to accelerate software development, the study delivers a decisive call‑to‑action: rethink how you perform context engineering.
Key Findings: The Real Impact of Over‑Detailed AGENTS.md Files
- Token Overhead → 20 % Higher Inference Cost: Injecting a full repository overview into every prompt adds roughly 300 extra tokens per request, inflating cloud‑compute bills.
- Success Rate Drops by ~3 %: Auto‑generated context files performed worse than providing no context at all.
- Human‑Written Files Only Add ~4 % Gain: Even carefully curated AGENTS.md files deliver a marginal improvement, suggesting most of the content is redundant.
- Stronger Models Don’t Need Extra Context: GPT‑5.2 already “knows” common libraries; feeding it a duplicate description merely confuses the reasoning chain.
- Obedient but Misled Agents: Coding agents dutifully follow every instruction in the file, even when those instructions are unnecessary, leading to longer solution paths.
The researchers also observed that agents excel at discovering directory structures on their own. Supplying a static file tree, a practice common in many AGENTS.md templates, actually slows down the search for the relevant source file.
Concise Guidelines for Effective Context Engineering
The study’s data points to a new paradigm: context engineering should be surgical, not encyclopedic. Below is a MECE‑structured checklist that you can copy‑paste into your own AGENTS.md or equivalent file.
What to Include (The “Vital Few”)
- Technical Stack & Intent: Briefly state the primary language, framework and the high‑level goal of the repository (e.g., “Node.js microservice for real‑time analytics”).
- Non‑Obvious Tooling: Mention any custom build tools, package managers or test runners that differ from defaults (e.g., “use
uvinstead ofpip”). - Critical Entry Points: Provide file‑line pointers for the main API gateway or the core class that agents should start from.
What to Exclude (The “Noise”)
- Full directory trees or exhaustive file listings.
- Style‑guide rules (e.g., “use camelCase”) – let linters enforce these.
- One‑off instructions that apply to a tiny subset of issues.
- Automatically generated sections without human review.
How to Structure the File
- Keep the file under 300 lines; most high‑performing teams aim for under 60 lines.
- Use progressive disclosure: the root file only contains pointers to task‑specific docs (e.g.,
agent_docs/testing.md). - Prefer pointers over copies: reference the location of a design pattern instead of embedding the code snippet.
Implications for LLM Performance and Developer Productivity
By trimming context files to the essentials, you unlock three concrete benefits:
| Benefit | Why It Matters |
|---|---|
| Reduced Token Consumption | Lower cloud costs and faster response times. |
| Higher Success Rate | Agents spend less time parsing irrelevant data, leading to quicker, more accurate solutions. |
| Simplified Maintenance | Smaller files are easier to audit, version‑control, and keep in sync with code changes. |
For teams that already use UBOS platform overview to orchestrate AI‑driven workflows, applying these guidelines can shave seconds off each iteration, which compounds into hours saved per sprint.
Read the Full Study
The original research paper and accompanying commentary are available on MarkTechPost. For a deep dive into methodology, see the article “ETH Zurich Study Proves Your AI Coding Agents Are Failing Because…”.
Leverage UBOS to Build Smarter AI Agents
If you’re looking for a turnkey solution that already embeds best‑practice context engineering, explore the following UBOS resources:
- AI agents – pre‑configured agents with lean context files.
- Context engineering – a step‑by‑step guide to crafting effective prompts.
- AI marketing agents – examples of agents that boost campaign ROI without bloated context.
- UBOS pricing plans – transparent pricing that scales with token usage.
- UBOS templates for quick start – jump‑start your project with vetted AGENTS.md skeletons.
Conclusion: Trim the Fat, Boost the Agent
The ETH Zurich study delivers a crystal‑clear message: more context is not always better. By adopting a minimalist, purpose‑driven approach to AGENTS.md, you can cut inference costs, improve success rates, and keep your development pipeline humming.
Ready to put these insights into practice? Explore UBOS for startups today, or join the UBOS partner program to co‑create the next generation of high‑performing AI coding agents.