State Machine Agent

How it is done

Model the interaction as a finite set of states (e.g. "greeting → collecting info → handling → done/escalate"), where each state allows only certain actions and a transition to the next state happens only when a condition holds. The model's freedom is confined within a single state; cross-state flow is rule-controlled. The system's behaviour space is thus finite and enumerable.

Trade-offs

Pros: predictable, testable, auditable, and able to enforce hard constraints like "never issue a refund before identity verification." Cons: states and transitions must be designed up front, and uncovered situations make it rigid; unsuitable for open-ended, exploratory tasks. It and a free ReAct loop are two extremes, and many systems pick per-subtask by risk.

Common misconceptions

• State-machine style does not mean no model — the model still works within each state; it just does not decide the overall flow.
• With incomplete state design, the agent gets stuck in a state or oscillates, so fallback transitions are needed.

Related terms

Sources

1. Finite-state machine — overview
2. LangGraph — low-level graph / state concepts