Three independent research groups tested whether context files — AGENTS.md, CLAUDE.md, the persistent instruction files that 60,000+ repositories now ship — actually help AI coding agents write better code.
They don't.
"Agents fail on implementation skill — feature design, pattern selection, exact wiring — not missing repository knowledge that a context file could supply."
— Khatri, 2026
The Evidence
Gloaguen et al. at ETH Zurich built AGENTbench — 438 tasks across 12 Python repos with real developer-written context files. Auto-generated files reduced success by 3%. Human-written ones improved it by 4%. Both increased cost by 20%. The formatting variables teams obsess over — file size, instruction position, architecture — produced no detectable effect across 1,650 controlled sessions.
Lulla et al. tested 10 repos and 124 changes. Runtime dropped 29%. Tokens dropped 17%. Completion rates held flat.
Khatri ran the most controlled study yet: 288 evaluated runs, two frontier agents (Claude Code, Codex CLI), 17 real tasks, gold-test evaluation from merged PRs. Three strategies — no file, always-on injection, selective wiki access. Correctness: statistically indistinguishable across all conditions. Claude scored 53.3%, 55.6%, 55.6%. Codex scored 58.8%, 56.9%, 52.9%. Omnibus permutation tests: p=1.00 and p=0.66. But wall-clock time dropped ~24%, and cache-creation tokens fell on 11 of 11 Claude tasks (p=0.012).
Three studies, one finding
Context files change what agents spend, not what they can do.
Faster completions. Fewer tokens. Same correctness.
The Adoption
OpenAI donated AGENTS.md to the Agentic AI Foundation under the Linux Foundation. Augment Code publishes guides on building effective context files. Twenty-eight tools list native support. The format has become infrastructure.
None of the vendor guidance mentions the null correctness result.
The pitch is always "make your agent understand your codebase." The research says agents already understand the codebase — they fail on what to do with that understanding. The file makes the route cheaper. It doesn't teach the agent to drive.
Whether that gap matters depends on what you were buying. If you adopted context files to cut token costs and speed up completions, the evidence says you got what you paid for. If you adopted them because a vendor guide said they'd make your agent write better code — the three groups that actually tested that claim found no support for it.
Sixty thousand repos. Three null results. The file is an economics lever, not a skill substitute. The industry knows which one sells.