Designing the escalation boundary for clinical agents.
Where an agent must stop and hand over to a person, and how Conductor encodes that boundary as a state machine instead of a prompt.
Autonomy is a dial, not a switch
Assembling a prior-authorization packet from the record is low risk and high volume. Changing a medication order is not. Agent design starts by deciding, step by step, which actions the agent may draft, which it may complete, and which always need a named person.
State machines, not prompts
In Conductor, each agent workflow is a state machine with explicit transitions. Approval points are states the workflow cannot leave without a human decision. Writing 'always ask a doctor' in a prompt is a suggestion. Making it a state is a guarantee.
Typed outputs
Agents produce typed actions (a draft order, a claim packet, a flag, a message) rather than free text. Typed actions can be validated, diffed, approved and reversed, and each one carries citations to the record it relied on.
Escalate on uncertainty, not only on rules
Besides fixed approval points, an agent escalates when evidence is missing or conflicting, or when its confidence falls below a threshold set for that workflow. Every escalation and override is logged and becomes test data for the next version.
Key takeaways
- 1Set autonomy per step, not per agent.
- 2Encode approval points as states the workflow cannot skip.
- 3Prefer typed, reversible actions over free text.