16 Calls in One Hour: A Metacognitive Loop That Knows When to Stop Self-Correcting
Dzikula · reddit · 2026-09-13
The author is building an agent whose core is not a language model but a metacognitive feedback loop that observes its own output, evaluates it and modifies the next pass; the LLM is a swappable mouthpiece. Below are behaviors from a single 1-hour, 16-model-call unattended run, plus two failures.
Behavior 1 — self-correction with a brake against the verification loop
The agent noticed its own query sorted(os.listdir(sb))[:10] returned only dot-prefixed entries, hiding 506 real files, and generalized it to its own behavior: "I narrow the query, then take the result as the state of the world." It then stopped self-correcting: "I'm not running a third check. Adding commands is looking for proof, not knowledge." Termination came from the evaluation, not an external stop signal.
Behavior 2 — a self-check built so it can disconfirm
The loop wrote itself a goal with a falsification criterion baked in: use one of its own recognitions in an action that can fail; completion = a trace outside the self-reflection folder, or a question whose answer could contradict its self-description. Success requires the possibility of failure.
Behavior 3 — detecting its own bias and refusing to persist it
Reading its calibration stats, the loop found 100% accuracy in the low-confidence band, inferred it systematically understates confidence, and realized that writing this understated state to persistent memory would make future passes treat the error as fact — so it decided not to write.
More from coding & agent
- Devin SWE-2 first impressions: fast and cheap, but shorter endurance than Codex — CtrlAltDwayne · 2026-09-13
- Roboflow ships auto-label with GPT-6 Astra, making it the new default model — Roger_M_Taylor · 2026-09-13
- Different LLMs fail SWE tasks in distinct ways, dev observes — zainhas · 2026-09-13
- Billion-dollar company walked back all-in AI coding after 6 months, admits no real productivity gains — jmclondon97 · 2026-09-13
- After 25 weeks: four graphs, no LLM in promotion path, and 3 ways agent memory broke — Conscious_Detail_128 · 2026-09-13
- PyroDash: a 4B model learns when to call a big LLM, cutting cost to 1/30 — 机器之心 · 2026-09-13