Bounded agent
Use a multi-step agent only inside a narrow objective, permission boundary, budget, and stop contract.
What this pattern means
Use a multi-step agent only inside a narrow objective, permission boundary, budget, and stop contract. Choose it because it is the least complex approach that satisfies the task—not because it uses the most capable model.
When it fits
- The work genuinely requires adaptive multi-step planning.
- Every tool and external side effect is bounded and observable.
- Monitoring, escalation, and deterministic stops are available.
When it does not fit
- A prompt, retrieval flow, or fixed tool sequence is sufficient.
- The agent can expand its own permissions or continue without a reliable stop.
Required controls
- Set task, time, cost, and action limits.
- Log steps and require approval for consequential transitions.
Autonomy limit
The maximum v1 autonomy is Bounded reversible action. A task-specific rule may impose a stricter cap.
Minimum evaluation
Run adversarial paths that test goal drift, prompt injection, permission escalation, loops, and failed stops.
Example
Reconcile duplicate records in a preselected queue with approval before merging.
Reassess bounded agent whenever the purpose, evidence, data class, decision owner, or external-action boundary changes.