We're hiringCareers at Moonlit Labs
Research

03 · Enterprise Intelligence

Oversight and trust in autonomous workflows

How much autonomy should an agent have, and what oversight makes that autonomy safe to grant?

Area
Enterprise Intelligence
Status
In progress
Models in scope
ClaudeGPTGrokDeepSeek

The question

Trust is earned in increments. This note looks at approval patterns, audit trails, and escalation paths, and at how different models behave when they should stop and ask.

Approach

  • Catalog oversight patterns used in agent deployments
  • Test how models handle uncertainty and out-of-scope requests
  • Study the cost of review against the cost of an unchecked mistake
  • Draft practical guidance for granting autonomy in stages

What we're watching

  • When does an agent ask for help, and when should it have?
  • What makes an audit trail useful rather than just present?
  • Where does oversight become theater?
NextFrom research to the road