FSM Design for LLM Agent Reliability

Existing AI agent frameworks give the LLM the role of an autonomous orchestrator that directly handles loop control, tool selection, and execution flow changes. This works fine for demos, but it has the drawback of making auditing, reproducibility, and formal verification difficult in tightly regulated environments such as finance (KYC/AML) or security (DevSecOps).

To address this, a new runtime model is presented that treats the LLM not as an autonomous agent, but as a bounded computational unit within a deterministic FSM (Finite State Machine).

Core Design Principles:

• Deterministic FSM Runtime: The LLM performs only a bounded number of computation steps, while the system itself directly controls transitions in the execution graph and the governance topology.
• ProjectionLayer (Evaluator Blindness): Instead of giving the LLM the full context, only a refined, target-specific projection is passed to it, preventing exposure of governance metadata and forming a security boundary.
• AST-based Condition Evaluation: Arbitrary code execution such as eval() or exec() is excluded, and conditions are evaluated through a constrained AST (Abstract Syntax Tree) engine to minimize the attack surface.
• Transition Entropy Monitoring: Rather than token count or latency, the variance and structural instability of execution paths are monitored to measure system stability.