Anthropic explains why 9 Claude Code agents burned 5M tokens mostly waiting on one GPU
audunlk · ghdev · 2026-10-13
A Claude Code user filed an issue about agents consuming massive tokens with little output, and Anthropic's diagnosis is instructive:
- Six 3D-scene agents each used 650K–730K tokens, over 5 million in total
- Most of the spend wasn't doing work but waiting: every time an agent woke to check whether its GPU render had finished, it re-read its entire long history
- Nine agents queued on a single GPU spent hours checking and re-checking, draining the usage limit
The fix:
- Cap concurrency at two agents, never more than the GPU can serve
- Run renders as one batch with no agent waiting on them; check results once when done
- Show results as soon as they exist, not reports about them
A textbook case of where agent token costs really come from: history re-reading inside polling loops.
More from coding & agent
- Claude cut a 2.6-hour video export to 20 minutes — but only after being told to rebuild the server — byhewar · 2026-10-23
- 3 flight-booking rules my AI agent checks that I'd never do by hand — Efistoffeles · 2026-10-23
- OpenJet claims Qwen3.8-27B coding runs 1.9x faster than llama.cpp on one RTX 3090 — JeremyCMorgan · 2026-10-23
- Building an autonomous SEO agent in Hermes with three skills: programmatic pages, Keywords Everywhere, DataForSEO — EXM7777 · 2026-10-23
- Building a sports betting AI agent with onchain execution and human approval by default: architecture and lessons — Dry-Comfortable-2678 · 2026-10-23
- How do you audit AI agent decisions in regulated workflows? A field survey — Ok_Way_1993 · 2026-10-23