ETH Zurich: auto-generated context files hurt coding agent performance in most cases

ETH Zurich researchers found that auto-generated context files like AGENTS.md worsened coding agent performance in five of eight test settings and increased inference costs by over 20%. The agents followed the instructions, but unnecessary requirements created cognitive load without improving problem-solving. Manually written context files only boosted success by about 4 percentage points. The study concludes AGENTS.md only pays off when hand-written to fill knowledge gaps missing from training data, like pointers to specific build systems or framework changes.

View full digest for February 18, 2026