Claude Code agents burned 9.45M tokens a day reading their own notes
WWEtrump · reddit · 2026-11-01
A solo dev running multiple Claude Code agents (engine, AI opponent, modeller, QA) to build a brick-style RTS in UE5 hit 9.45M tokens in one day — most of it agents reading their own notes: whole source files when 40 lines would do, logs ballooning from 1M to 2.5M chars, and a handoff file that grew to 373,789 chars (93k tokens per full read). His fixes:
- A one-page Docs/BRIEF.md (1,350 tokens) every agent reads first
- A 6,000-token reading budget per role before work starts
- A compactor role running compactdocs.py that archives finished work byte-for-byte into Docs/archive/ with pointers (295k chars moved on first roll)
- Three CLAUDE.md rules: never read whole files; edits, never full rewrites (rewrites resend files and can silently drop lines); one deliverable per run, stop near 250k tokens and hand off a short state
Hardest lesson: a resumed agent isn't a continuing worker but a very expensive reader — one resumed cinematics agent cost 450k tokens per swap. Docs on long projects are a cost, not just context; log file sizes and tokens per run before blaming the model.
More from coding & agent
- Hands-on tutorial trains a decision model on Qwen3-0.6B from scratch — dl_weekly · 2026-11-01
- ui-ux-pro-max and 6 more repos giving AI coding tools real design taste — GCWebDesigner · 2026-11-01
- Hex ships stable release: local voice dictation for talking to AI agents on Mac, Windows, Linux — madeofAjala · 2026-11-01
- Personal AI agent race heats up — this prompt migrates your entire setup between agents — EXM7777 · 2026-11-01
- Nous Research's Hermes Agent launches mobile beta on iOS and Android via TestFlight — testingcatalog · 2026-11-01
- How to Build a Production AI Agent: The 7 Layers Between the Model and the Result — blaizedsouza · 2026-11-01