Treat each skill/SOP as a hypothesis with a success posterior, then patch/split/compress/retire/explore (Bayesian-Agent)

arXiv 2606.08348·high signal

Bayesian-Agent reframes prompt and skill evolution as maintaining a feature-conditioned posterior over whether a frozen model will succeed with a given skill in a given context, updated from verified trajectory evidence. Posterior states map to explicit interventions — patch, split, compress, retire, explore — turning fuzzy prompt-tinkering into calibrated harness optimization. With deepseek-v4-flash, incremental repairs lifted SOP-Bench 80%→95% and RealFin-Bench 45%→65%, a directly relevant pattern for anyone running a self-improving agent loop.