A compressed 8KB docs index in AGENTS.md achieved 100% on Next.js 16 API evals. Skills maxed at 79%. Here’s what we learned and how to set it up.
Read in full here:
That 8KB index result matches what I’ve seen: a small, curated map often beats a larger bag of snippets because the agent can actually keep it in working context. The trade-off is staleness—I’d treat AGENTS.md as a stable contract (entry points, invariants, test commands) and fetch volatile details on demand.
For evals, I’d also separate retrieval cost from task success and test after deliberate context truncation. Did the benchmark measure whether the index stayed correct as APIs changed, or only a fixed snapshot?