Human-in-the-loop supervision keeps people able to inspect an agent’s actions, catch errors, and redirect its work. More capable agents can require more oversight: file access, edits, and command execution all create review demands. section 5.2.3
Match oversight to risk rather than approving every action reflexively:
- Use selective trust: reduce friction for routine work and supervise novel or destructive actions closely. section 5.2.3
- Watch for approval fatigue; rubber-stamping can turn a safety check into a source of risk. section 5.2.3
- When agents run unattended, replace continuous human feedback with validation checks, anomaly detection, audit logs, circuit breakers, and sampled human review. section 5.3.4
Removing people from the loop can let agents process work while nobody is watching, but errors and drift may remain unnoticed until downstream consequences appear. Traceable records of agent context, actions, tools, and supervision make later debugging and verification possible. section 5.3.4 section 5.5.2
Approval cost
An agent that saves thirty minutes of coding but takes twenty minutes to supervise offers only a modest net gain; rubber-stamping a destructive action can erase that gain. section 5.2.3



























