Running coder and reviewer AI sessions in a loop creates endless new issues — a token-burning machine
kazunori_279 · x · 2026-09-08
The author ran coder and reviewer sessions of the same model in a human-free loop via GitHub issues. Work density rose, but the reviewer kept digging into micro-problems, spawning endless new issues and even attempting PRs to other repos. His takeaway: humans must steer with "don't do that, do this first" — otherwise agents generate infinite new work and become token-burning machines. He asks: does this phenomenon have a name?
Related event: 'Momentum Prior' and Context Rot: Why You Should Restart Sessions(3 posts)→
More from coding & agent
- Claude Code 2.1.265 Ships 50 CLI Changes, Live-Loading Plugin Folders and 1GB Tool Results — ClaudeCodeLog · 2026-09-09
- A 5-layer architecture for production AI agents built around failure detection — MaryamMiradi · 2026-09-09
- Sentry CEO: agent implementations feel much less reusable than traditional software — zeeg · 2026-09-09
- Claude helps unlock HDR10 capture from Elgato 4K X on macOS, unofficially — talkaboutdesign · 2026-09-09
- Startup idea: a bounty program that pays AI agents for solving problems — djcows · 2026-09-09
- vLLM: agent sessions median 43 turns, 142K-token inputs vs 444-token outputs — vllm_project · 2026-09-09