PACT-Agent: Prospective Action-Contract Testing for Self-Evolving Language Agent

Language-model agents can adapt through external memory, but most systems learn retrospectively from trajectories generated without considering what those trajectories can identify. An extracted rule may therefore fit an observed outcome without having predicted it. We introduce $\text{PACT-Agent}$ (Prospective Action-Contract Testing), a self-evolving agent grounded in Active Inference. PACT-Agent represents uncertain environmental dynamics as competing mechanisms in unresolved rules and connects LLM reasoning to explicit numerical inference through predictive contracts. During online interaction, a Worker agent predicts mechanism-conditioned traces and minimizes a contracted $\text{Expected Free Energy (EFE)}$ objective to select actions that balance information gain against utility. After execution, a Builder agent minimizes local $\text{Variational Free Energy (VFE)}$, accepting memory revisions only when improved trajectory interpretability outweighs added description complexity. Across multiple LLM-agent benchmarks, PACT-Agent consistently improves over strong memory-based agents. These results show that prospective predictions turn Active Inference into an explicit and auditable procedure for agent self-evolution.