Self-learning agent loop takes Honor of Kings win rate from 0% to 60% in three rounds, no retraining
kimmonismus · x · 2026-11-17
A Singapore AI lab (@agnesaisapiens) built a self-learning loop where a coach model reviews finished games, diagnoses what went wrong, and rewrites the team's playbook — changes only persist if they pass tests. No model retraining involved.
- Bare team lost all 192 initial games (0% win rate)
- After three loop rounds: 60% win rate (75/124) vs the built-in bot; the chart shows later rounds reaching 75%/90%/95%
- Key ablation: with only results and digests as input, the coach produced 0 wins across 2,318 games; given the full game replay and room to think, it hit 60% in three rounds
kimmonismus argues this addresses agents' most annoying failure mode today — repeating the same mistakes until someone manually patches the prompt. Systems that review their own work and keep only effective changes are what he wants to see more of.
More from coding & agent
- A practical CI prompt: make your tests and builds use every CPU core on the runner — DanielLockyer · 2026-11-17
- The boring work behind successful AI agents: data and process cleanup decide outcomes — bigdata · 2026-11-17
- J.Doe enters private beta: proactive decision cards instead of empty prompt boxes — umesh_ai · 2026-11-17
- Building an Obsidian music notation library and practice plugin with Fable 5.2 — dSebastien · 2026-11-17
- Mistral makes Vibe free in Europe for 3 months, claims coding edge over GLM and DeepSeek — qtnx_ · 2026-11-17
- Cheap tokens, expensive agent: Flash hits $11.97/task vs Luna's $0.05 in Databricks AI Index — krishnan · 2026-11-17