Definition
Plain language
A system that turns an AI agent's plan into a formal object and checks whether the plan even makes sense before the agent runs it.
As stated in the literature
A framework that encodes an agent workflow as a typed graph in Lean4 and applies layered static verification — structural well-formedness, Hoare-style precondition/postcondition contracts, and graph-level information-flow checks — plus runtime trajectory checking and step-localized repair.
Why it matters: It catches flawed agent plans up front rather than after they fail halfway through, saving the cost and risk of running a doomed workflow.
For example, before an agent starts a multi-step task, Lean4Agent checks that step three actually produces the information step four expects, catching a broken plan before any work is wasted.
Heard on the show
“The paper is — full title — "Lean4Agent: Formal Modeling and Verification for Agent Workflow and Trajectory.”Episode 122 — When Your Coding Agent Lies About the Fix: Verifying the Plan Before the Model Runs