[Bug] Agents consume excessive tokens without producing results
Status Open
Reported on v2.1.296
Maintainer reply None cached
Activity 3 comments · opened Oct 12, 2026
Bug Description
The agents consumed a ton of tokens without doing any actual work, they just ingested without any results. Please refund me
Environment Info
- Platform: win32
- Terminal: windows-terminal
- Version: 2.1.296
- Feedback ID: 29d29a14-ea27-4f5a-9b63-7832f0e8ebc4
Did they do any work? Yes, but too little for what they cost.
- The work: each agent wrote its 3D scene from scratch and rendered test frames. Six scenes produced frames that clear your bar; you'll see them on the board.
- The cost: each agent used roughly 650,000–730,000 tokens, over 5 million in all. Most of that went on waiting for the GPU. Every time an agent woke to check whether its render had finished, it re-read its whole long history, and that re-reading is what costs. With nine agents queued on one GPU, they spent hours checking and re-checking, and that is what ran out your usage limit.
From now on:
- At most two agents at a time, and never more than the GPU can actually serve.
- Renders run as one batch, with no agent waiting on them. I check the results once when the batch finishes.
- You see results as soon as they exist, not reports about them.
Next I'll put the six finished scenes on the board, so you can judge them yourself before anything else is built.
3 Comments
Found 3 possible duplicate issues:
This issue will be automatically closed as a duplicate in 3 days.
🤖 Generated with Claude Code
If you take the refund request further, the strongest support is numbers from your own transcripts rather than the agent's summary. Each agent's transcript (
%USERPROFILE%\.claude\projects\<project>\*.jsonl, with subagents in their own files) records every model call's input, cache reads, cache writes and output. The pattern the summary describes, agents repeatedly waking to check a render and re-reading their whole history each time, shows up there as many calls with a large context and almost no output, each following asleepor a background-output check. If the checks were spaced more than the cache lifetime apart, they also show up as full cache rewrites.A free script that totals exactly that per transcript (disclosure: I wrote it):
npx github:lowkey-studio/quotatrace --only claude --since 2026-10-11. It reports calls after a sleep or background check and their token cost, big-context calls that produced under 300 output tokens, cache rebuilt after a pause, subagent share, and the heaviest transcripts. The output is counts only, no prompts or paths, so you can attach it here or to a support request. https://github.com/lowkey-studio/quotatrace (there's also an optional paid review if you want someone to read it and draft the request, but the numbers alone are free).When tokens burn without results, the usual suspects from auditing agent usage logs: retry chains (each failed attempt re-sends the working context, and failures repeat when the fix needs a human — a flaky check, a wrong path, an auth problem the agent can't self-heal), and duplicate usage records after a crash or resume. Both drain limits while the visible output stays thin.
You can see it directly in your own session files:
pip install agentmeasure→agentmeasure checkscans local Claude Code logs for retry chains, consecutive-failure runs and duplicate records — local only, nothing uploaded. If the report says e.g. "3 executions carry 1 logical operation" over a failing command, that's the token sink made visible, and usually points at the one thing worth fixing once.