skip to content
Skylar Payne

An unconstrained agent loop is easy to demo and miserable to operate.

An agent can analyze, search, draft, call tools, ask questions, revise, escalate, and start over. The flexibility feels powerful until it repeats a step, skips a required check, or jumps from vague analysis straight into a risky action.

By then, “the agent decided” is not an explanation. It is a bug report with worse nouns.

Model the process as a state graph.

Give the process named states: intake, classify, retrieve context, draft, validate, request approval, execute, done, failed. Define which transitions are allowed. Store the current state and the reason for each transition. Make loops intentional and bounded.

The model can still help choose or execute steps. It operates inside machinery the rest of the product can inspect.

In practice, you get:

  • progress is visible
  • impossible transitions are rejected
  • retries have limits
  • approval gates have a real place to live
  • failures can resume from a known state
  • evals can target individual transitions
  • humans can understand what happened after the fact

There is the line: magic while-loop on one side, software you can operate on the other.

Once the agent touches customers, tools, or production state, the graph stops being an architecture diagram and starts being insurance. A state graph gives agents the boring machinery they need: locality, traces, replay, and approval.

It also keeps AI reliability work grounded. Instead of asking whether the whole agent is good, ask whether a transition should have happened, whether a state had enough evidence, and whether the next action required a human.

If the agent matters, make the states explicit. Let the model do bounded work inside the graph, not invent the graph at runtime.

Part of the Effective AI Engineering series.

Source: adapted from Mirascope’s “Graph-Based Agent Workflows”, MIT licensed.