Research
Credence: Principled Bayesian Agent Architecture
A framework for LLM tool-using agents based on decision theory rather than prompt engineering. The core insight: whether to query a tool is a decision problem, solvable with Value of Information. The core finding: optimising for accuracy alone is counterproductive when queries have costs.
Credence maintains Beta posteriors over tool reliability and computes expected value of information before each query. On a benchmark task, LangChain ReAct scored -8 (63.7% accuracy) while Credence scored +112 (59.6% accuracy). The agent that answered fewer questions correctly won decisively — because it knew which questions were worth asking.
This connects directly to open problems in agent alignment: an agent that maximises a coherent utility function under uncertainty is more controllable than one that maximises a proxy metric. The accuracy paradox is a toy demonstration of Goodhart’s Law in agent architectures.
- GitHub
- Agentic AI Is Neither Intelligent Nor an Agent — the benchmark result
- How Decision Theory Cuts Your AI Agent’s API Bill in Half — the implementation
- The Bitter Lesson Has No Utility Function — why scale alone doesn’t solve the decision problem
arXiv preprint: forthcoming.
Published Work
Freeman, G. & Smith, J.Q. (2011). Bayesian MAP model selection of chain event graphs. Journal of Multivariate Analysis, 102(7), 1152–1165. Chain Event Graphs — a class of graphical models more expressive than Bayesian networks for asymmetric problems — and Bayesian model selection over them. Foundational paper in the CEG literature; subsequent work on dynamic CEGs, causal algebras, and software builds on these methods.
Freeman, G. & Smith, J.Q. (2011). A Bayesian approach to event trees. Bayesian Analysis, 6(4). Non-stationary sequential models with conjugate Bayesian updates. The exponential forgetting mechanism for regime change prefigures methods now used in adaptive agent architectures.