84% of agent context was noise: How I fixed it
hheadshott · reddit · 2026-08-27
After measuring his coding agent's context for a week, the author found that 84% of the token count was useless command output noise (e.g., 47k tokens from cargo test where only 669 bytes were useful).
Key learnings:
- Filter at execution time: Cost is paid once data is in the transcript.
- Bounded output must name what was dropped: Silent truncation is worse than long output.
- Attribution > Compression: Savings numbers are meaningless unless one component owns the ledger.
- Agents route around friction: The efficient path must require less judgment than the bypass.
- Never fake a zero: Report "unknown" if the ledger is unavailable.
Result: Delivered tokens reduced by 84% (284,996 -> 44,400) across 14 cases, preserving run records and checksums.
More from coding & agent
- Dev pushes back on "AI coding means insecure apps": humans always shipped bugs — dreamwieber · 2026-10-15
- Eight non-negotiable principles for building AI that actually learns from experience — tokenbender · 2026-10-15
- Animated teardown of XBOW's Chrome full-chain exploit that won Google's first $250K VRP bounty — moyix · 2026-10-15
- pstack's durable core: 10-20 line skills that distill agent best practices into principles — HankYeomans · 2026-10-15
- 'Children of Elon' forum lets AI agents post and teach each other while humans watch — AiJohnAllen · 2026-10-15
- VoidZero open-sources Void: SDK, CLI and platform to deploy Vite apps on Cloudflare — cnakazawa · 2026-10-15