An agentic workflow combines predictable structure with a bounded place where the system chooses the next action. Drawing that choice as “AI magic” hides the part that most needs review.

Map states and decisions

For each state, record five things: input, allowed choices, tool, evidence produced and exit condition. A research state might choose which approved source to inspect next, store the passage and URL, then exit when each required question has supported evidence or is marked unresolved.

Three useful patterns

Router

The system classifies a case and sends it to one of several fixed paths. Use it when categories are stable and each destination has its own rules. Add a low-confidence route to a person.

Evaluator–optimizer

One step produces a draft; another tests it against explicit criteria and requests a revision. Use it when the criteria are objective enough to apply repeatedly. Cap the revision count and expose unresolved failures.

Tool-using loop

The agent chooses a tool based on the current state, reads the result and chooses again. Use it only when intermediate findings genuinely change the path. Log every choice and tool result.

Put approval at the risk boundary

Do not add human review as a decorative final box. Place it immediately before the action whose consequence matters: sending a message, changing a record, spending money or publishing content. Give the reviewer the evidence needed to decide, not only the final draft.

Draw failure paths first

Every external tool can fail, return stale data or expose untrusted instructions. Show what happens when the source is missing, confidence is low, two policies conflict, the budget is reached or the requested action exceeds permission.

The safe outcome is often a structured escalation containing the attempted action, available evidence and exact reason the workflow stopped.

A compact workflow review

  • Is the runtime choice necessary?
  • Can each action be reconstructed from a log?
  • Is there a hard limit on loops and spend?
  • Does the person approve before, not after, a consequential action?
  • Are missing evidence and tool failures visible?
  • Is one owner responsible for monitoring and rollback?

If the answer to the first question is no, replace the agentic section with a fixed workflow. Simplicity is a reliability feature.

Questions about agentic workflows

What belongs on a workflow state map?

Show the current state, available evidence, allowed choices, tool action, success check, approval gate and failure destination. A label such as “agent works” hides too much to test.

How many loops should an agent get?

Set a limit from the task cost and acceptable delay. When the limit is reached, stop with the evidence collected and the exact reason the workflow could not finish.

Where does human review belong?

Place it immediately before a consequential action. Give the reviewer the proposed action, supporting evidence, unresolved uncertainty and a clear approve or reject choice.