loading
Loading.loading
Loading.Usually not, done the obvious way. A 2026 ETH Zurich study (Evaluating AGENTS.md, arXiv 2602.11988) found that adding repository context files tends to reduce coding-agent task success by about 3% and raise inference cost over 20% versus no repo context. Only short, human-written files that state minimal constraints helped. The win is serving the right context per task, not a longer file.
Updated
Attention is finite. A big project overview buries the few lines that matter for this task under the ones that don't ('lost in the middle'). The agent reads more, wanders more, and pays per token, so you get lower success at higher cost.
The study's own result is that minimal human-written constraints help a little; big auto-generated dumps hurt. A good AGENTS.md is short, current, and about constraints, not architecture tourism.
The durable fix is retrieval that serves currently-correct, scoped, fresh context per task instead of a static file that ages the day it ships. trovex does exactly this and cuts ~60% of the tokens per lookup. tsukumo builds that context layer with your team.
or have us build it — same capability, the other door