Coding agents tend to copy existing repository patterns even when new task instructions conflict. Engineers should state intent explicitly and not rely on instructions to override precedent in the code.
Key takeaways · AI-distilled
Across 5,625 trials with four coding agents, the authors found the agents share the same default implementation choices but followed repository precedent over those defaults in 359 of 360 conflicting trials.
When a new task's intent conflicted with existing code, written priority rules beat code examples alone, with 59 of 60 joint successes versus 38 of 60 for two agents on two implementation pairs.
In four open-source repositories, a priority rule loaded by the agent harnessThe scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.Full definition → raised joint success from 0 of 120 to 117 of 120, per the abstract. The authors argue for recording design rationale and priorities alongside the code.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
articleStale Instruction Files Override the Code in Coding Agents: A One-Sentence FixXiaonan Xu, Wenjing Wu
articleWhat Open-Source Projects Tell Their Coding Agents: A Census of AGENTS.md and CLAUDE.mdPawel Zarzycki
articleNot All Agents Are Equal: Code Quality and Post-Merge Maintenance Across Five Autonomous Coding Agents in the WildObada Kraishan