Running >1 autonomous agent: doc bloat, review bottlenecks, and why "don't edit" fails

menhguin · x · 2026-10-21

A developer shares hard-won lessons from running more than one autonomous agent:

Docs explode: a few KB single-agent docs ballooned to hundreds of KB within days — a serious cost/context issue even at 1MM context.
Manual review breaks: comprehensive multi-agent changes are nearly impossible to review; he's testing a cleanup agent that fixes obvious errors, sweeps open tasks, and reads frontier-lab multi-agent reports for top-priority fixes.
"Don't edit" is hopeless: assume agents will eventually edit; instead add release valves dictating how (append at bottom, not inline).
Context retrieval > rules: instruction files drift and need human approval, so keep instructions minimal and rely on immutable primary sources (data, chatlogs, bookmarks) with frontmatter/embeddings.
Plain filesystems don't scale — with 10 agents you need a fuzzy-intent tool for locating instructions.
Reviewer agents with separate personas are unnecessary; reviewer skills/crons suffice.
The minimum for async agents: detailed user priorities/context + a heartbeat trigger, starting from a system the user actually gives feedback on.
Opus is currently much better at handling autonomy than Astra.

More from coding & agent