Problem
Agents without limits fall into costly loops, call tools too often and sometimes take actions nobody approved.
Solution
Each tool is classified by reversibility. A supervisor controls the run budget and pauses for human approval before irreversible writes.
When to use
- →Agents with access to real systems
- →Financial actions, sending messages, deletions
- →Teams starting to grant agents autonomy
## Trade-offs
- +Predictable cost per run
- +Predictable failure instead of silent failure
- +Autonomy grows with trust
- −Approvals add latency
- −Requires designing the failure path