AGENTS.md Files Don't Move Coding Agent Correctness: A 288-Run Ablation

AGENTS.md and CLAUDE.md files are the standard way teams steer coding agents. New research asks the question most of us have never tested: does the context actually change what the agent ships? The answer, from a controlled ablation of 288 evaluated runs across Claude Code and Codex on 17 real tasks from 3 repositories, is that injection strategy does not measurably move correctness on either agent, bounded to under 10-15 percentage points via equivalence testing.

The paper also explains why, and the why is more interesting than the null.

What the study did#

The researcher, Prakhar Khatri, built a SWE-bench-style harness with three context-injection strategies applied to the same tasks:

• none: the AGENTS.md is removed from the workspace entirely
• always_on: the full file is injected into the system prompt every turn
• selective: topic-organized wiki files are placed in the workspace and the agent reads them on demand

Each task ran under all three strategies with 3 independent repeats, on Claude Code (claude-sonnet-4-6) and Codex CLI (gpt-5.5). Tasks came from merged pull requests in three Python repositories: pdm (477-word context file), firebase-admin-python (1,236 words, rated "Excellent" on the study's rubric), and opshin (248 words). Correctness was scored by hidden gold tests extracted from the merged PRs, and runs executed on an egress-locked pod with GitHub DNS blackholed so agents could not read gold solutions.

Pass rates by strategy:

Omnibus permutation tests found no strategy effect (p = 1.00 for Claude, p = 0.66 for Codex). Equivalence testing bounds every pairwise difference to under 10pp for Claude and under 15pp for Codex.

Why context does not rescue these tasks#

The mechanism is the durable finding. The author triaged near-miss failures (1-4 failing gold tests), the exact tasks where one extra fact could flip a pass. None of them were missing-knowledge failures:

• a union-expansion optimization pass built correctly but with a correctness bug
• reactive retry implemented where the fix required proactive auth-token refresh
• a validator check the agent knew about but miswired
• type-narrowing that needed deep type-system reasoning

The failure mode is implementation skill: feature design, pattern selection, exact wiring. A context file cannot supply that. A manipulation probe hammered the point home: the two convention-closest near-miss tasks were re-run under all strategies on both agents, 36 cells total, and the real, unmodified AGENTS.md never converted a near-miss to a pass on either agent (Codex failed 18/18 regardless of strategy).

Context still changes how agents work#

The correctness null does not mean context is inert. Two process-level signals survived:

• On opshin, whose AGENTS.md warns that the full test suite takes over 20 minutes, blind full-suite pytest runs dropped monotonically with context dose: none 3.67, always_on 2.44, selective 1.67 per cell. Wall-clock time fell about 24% (2,689s vs 2,066s vs 2,032s). The file changed test strategy, not outcomes.
• For Claude, selective injection cut cache-creation tokens on 11 of 11 tasks (Holm-corrected p = 0.012), a mechanical artifact of a shorter system prompt.

Why prior studies disagreed#

The study offers a clean explanation for the contradictory prior results: borderline task difficulty is agent-specific. Across the 15 shared tasks, per-task pass rates correlate at Spearman rho = 0.75, but roughly 40% of tasks sit in different difficulty bands per agent (borderline for one, floor or ceiling for the other). A task set calibrated on Codex is mostly floor/ceiling for Claude, where no manipulation can register. Single-agent studies draw tasks from different agents' informative bands and reach different conclusions without any contradiction in the underlying behavior.

A power analysis makes the practical stakes clear: at 15-17 tasks with 3 repeats, even a 30pp effect is only caught 57% of the time, and detecting a 10pp effect at 80% power needs roughly 120-200 tasks. Most "context files work" or "context files are useless" takes floating around are built on studies that structurally cannot detect the effect they claim.

My take#

This is the most rigorous AGENTS.md study yet, and its honest reading is not "delete your context files". It is:

Context files are behavior steering, not capability injection. If you want them to change correctness, write them for the failure modes that actually gate tasks: worked examples, task decomposition, wiring patterns. The paper's own recommendation is that effort spent on generic convention prose may pay off less than task decomposition, tooling, or example-driven prompting, which lines up with why example-rich context wins in our context engineering guide and in our coverage of Claude's context engineering rules.

Process effects are real and valuable. We run this site's fleet on AGENTS.md files with verification gates and commit rules, and the opshin result is exactly the mechanism we rely on: files that tell agents when and how to run expensive checks change behavior in measurable ways even when pass/fail does not move. That is the argument behind keeping context lean, which we covered in agent context reduction patterns and running a fleet of Claude agents.

Calibrate per agent. The rho = 0.75 finding is a methodological warning for everyone building agent workflows: what is borderline for Claude Code may be trivial for Codex and vice versa. Our comparison of repository context for coding agents and the Claude Code skills primer both touch the same trap from different angles: guidance that changes one agent's behavior does not transfer to another.

Caveats worth stating plainly: this is 3 Python repositories, naturalistic style-guide-type files, and a snapshot of two model versions as of late July 2026. Purpose-built, task-specific context remains an open question. But the burden of proof just shifted: teams claiming AGENTS.md files improve agent correctness should now have to show it at scale, not assume it.

Continue Reading#

• The Text Is a Control Surface - the boundary of this null: decorative prose pays nothing, but actionable interface text and re-grounding after lossy stages carry measurable weight
• The Context Engineering Guide: Getting Claude Code to Actually Read Your Repo
• Claude 5 Context Engineering Rules: What Changed in the System Prompt Era
• Managing a Fleet of Claude Agents
• Agent Context Reduction: Cutting Tokens Without Losing Behavior
• What Are Claude Code Skills? A Beginner's Guide

Sources#

• Do Context Files Help Coding Agents? A Two-Agent Ablation Study on Real Repositories - arXiv:2607.27250 (abstract)
• Full paper HTML - arXiv:2607.27250v1
• Evaluating AGENTS.md: Are Repository-Level Context Files Helpful for Coding Agents? - Gloaguen et al., arXiv:2602.11988 (prior work cited by the study)