What would it take to build an agent whose behaviour is derived from a few fundamentals the way physics is derived from conservation laws? Three types, four axioms, and a refusal to add anything else.
Credence-proxy sits between an agent and its LLM providers, learns which model is good for which category, and routes accordingly. On an OpenClaw benchmark it cut cost by 96% and latency by 52% while raising quality by 1.24 points. The mechanism is one equation.
Each prompting technique helps. Reasoning traces, strategy guidance, cross-question history --- each one improves accuracy and score. None of them closes the gap with a Bayesian agent that does not use language at all. The ceiling exists because descriptions of calculations are not calculations.
Most agents are given a fixed set of decision rules. Credence's second tier generates candidate rules from sensor features, scores them by complexity, and lets the posterior decide which structures are worth keeping. This is program synthesis as Bayesian inference.
A Bayesian decision-theoretic agent scores lower on accuracy than every LLM variant it competes against --- and beats the best of them by 120 points. The explanation requires thinking about something that LLM benchmarks typically refuse to think about.
Every RL agent that has played a text adventure has tried to take the lantern fifty times in a row. The fix is not better exploration heuristics. The fix is representing state properly.
A Bayesian agent plays text adventures with four information sources and a VOI gate on every query. The LLM recommends actions. The mathematics decides whether to listen.
I wrote about decision theory fading from AI. Hacker News said I was annoyed at Rich Sutton's Bitter Lesson. I wasn't. But the misreading proves the point.
Someone asked why decision theory stopped being widely used in AI. The answer involves ImageNet, academic departments, and the seductive power of not having to specify your objectives.
Most LLM agents use tools like a tourist uses a phrasebook. A few lines of probability theory can fix it.
I built a Bayesian agent and set it against LangChain on a tool-use benchmark. LangChain got more answers right and still lost — by 120 points.
Part 2 of the Bayesian agent series. We confront the question of what should be designed versus what should be allowed to emerge, and discover that it's agents all the way up and all the way down.
A visual demonstration of Bayesian inference and the exploration-exploitation tradeoff. An agent navigates a grid world, learning which foods are safe through exact conjugate updates.