Human Approval for AI Agents: State Machines

Khourshid describes keeping the overall workflow deterministic while allowing an agent flexibility within each step. That gives teams control over which steps can happen and in what order, while still letting a model handle open-ended tasks.

He uses an email drafter as an example: the agent drafts a message, but the state machine prevents sending it before user approval. He says prompts alone cannot guarantee business rules. Mapping traces to states and transitions can also show where an agent gets stuck or spends unexpected time.

Khourshid says state machines can be implemented in any language, from a switch statement to a function that returns the next state. The workflow repeatedly reads events, chooses a state, runs its effects, and continues until the process is done. He sees actor models as a way to coordinate agents and state machines as a way to define each agent’s behavior.

Because model output may be unexpected, Khourshid recommends using guards such as schema checks to determine whether an output is valid. The workflow can then retry the agent call or move to an error state for mitigation. At Stately.ai, an earlier system generated three candidate state machines, checked them, and looped back when they were invalid.

Khourshid says simple input-output tasks may not need a state machine, while some production machines reach 200 states. He points to LangGraph as a graph library for defining agent logic with state machines. Stately.ai offers a visual editor at states.new.