The agent doom loop isn't the model being dumb — it's the transcript working against you
RunAI_Coder · reddit · 2026-09-01
A widely discussed Reddit post argues that agents stuck retrying the same failed fix aren't dumb: since the model is stateless and re-reads the full transcript every turn, after four failed attempts the dominant pattern in context IS the failed attempt. A next-token predictor treats that history as what this session does, not as evidence the approach is wrong.
Key evidence: a SWE-bench trajectory study found agents in failed runs had located the correct file 72–81% of the time — finding the spot was never the problem; letting go of the hypothesis was.
Fixes live in the harness, not the prompt:
- A hard turns/token budget so a stuck run stops instead of politely burning money
- Fingerprinting attempted diffs so near-identical retries trip a forced "list three untested hypotheses"
- The nuclear option: clearing the window entirely, with learned constraints re-injected in a few dozen tokens in the new opening prompt
The author closes by asking for a repetition detector that doesn't false-positive on legitimate retries (flaky tests, rate limits).
More from coding & agent
- DIY visual diff tool using GitHub Artifacts and pure JavaScript to save costs — zeeg · 2026-09-01
- Dev shares a dirt-cheap approach to visual diffs — zeeg · 2026-09-01
- Grok Bot automates Shopify updates and supplier coordination — billyjhowell · 2026-09-01
- Grok Bot automates lost deal analysis by mining call and email threads — lennysan · 2026-09-01
- Design pattern: immutable agent artifact revisions behind a stable review URL — RocketSeven · 2026-09-01
- Building a long-term memory benchmark for agents: what to add? — True_Mongoose_7073 · 2026-09-01