**Bug Description** The numbers: - Two sessions, 23 Opus agents, 13,033 model turns. One was the evening's 0.2.3 engine build, the other the overnight loop. - 5.57 billion tokens re-read, about 1.6 million written. Almost none of the usage was actual work. - Agents' conversations grew to 400–970k tokens and were never restarted. Every turn re-sends the whole conversation. 72% of the night loop's re-reads were context above 150k. - Several agents ran 5–7.5 hours across many tasks. The

[Bug] Long-running agent sessions re-read unbounded context, causing excessive token usage

Status Open

Reported on v2.1.299

Maintainer reply None cached

Activity 0 comments · opened Oct 14, 2026

Bug Description
The numbers:

Two sessions, 23 Opus agents, 13,033 model turns. One was the evening's 0.2.3 engine build, the other the overnight loop.
5.57 billion tokens re-read, about 1.6 million written. Almost none of the usage was actual work.
Agents' conversations grew to 400–970k tokens and were never restarted. Every turn re-sends the whole conversation. 72% of the night

loop's re-reads were context above 150k.

Several agents ran 5–7.5 hours across many tasks. The main loop ran 10 hours straight.
About 13% was status checks. Agents kept checking on builds and benchmarks, and each check re-read around 460k tokens.
570 full-size images stayed in context and were re-read on every later turn.
Everything ran on Opus, including benchmark babysitting.

A fresh agent per task, capped near 150k, would have cut the total by about 60%.

I saved these as standing rules for every future session:

one task per agent, restarted around 150k;
no status-check turns while waiting;
downsized images reviewed in short-lived agents;
Sonnet for mechanical work;
a stated token budget on long or overnight runs, with the spend reported at each check-in.

Environment Info