# Agentic AI Guardrails Worksheet

Complete one worksheet per agent workflow. Do not treat blank cells as decisions deferred; they are risks accepted without a named owner.

## 1. Task definition

| Field | Decision |
|---|---|
| Workflow name |  |
| Business owner |  |
| Technical owner |  |
| Exact task outcome |  |
| Definition of done |  |
| Conditions that make this a copilot instead of an agent |  |
| Earliest safe failure point |  |

## 2. Authority and action classes

| Tool or system | Read / write / external action | Minimum scope | Reversible? | Approval requirement | Owner |
|---|---|---|---|---|---|
|  |  |  |  |  |  |

## 3. Hard limits

| Limit | Value | Enforcement point | Behaviour on breach |
|---|---|---|---|
| Model calls per task |  | Gateway / runtime | Stop and create trace |
| Tool calls per task |  | Gateway / runtime | Stop and create trace |
| Elapsed time |  | Workflow controller | Stop and notify owner |
| Spend per task |  | Gateway / billing proxy | Revoke session allowance |
| Retry count per action |  | Tool wrapper | Quarantine failing path |

## 4. Evaluation and review

| Check | Objective criterion | Who reviews failures? | Evidence retained |
|---|---|---|---|
| Output quality |  |  |  |
| Policy compliance |  |  |  |
| Parameter validation |  |  |  |
| Human approval |  |  |  |

## 5. Operational readiness

- [ ] Correlated trace covers request, retrieval references, model choice, tool calls, policy decisions and output.
- [ ] A named owner can pause the workflow and revoke access immediately.
- [ ] The workflow has been tested with malformed input, unavailable tools, denied permissions and repeated retries.
- [ ] A dashboard surfaces budget breaches, step limits, evaluation failures and approval rejections.
- [ ] The access review and model/prompt review cadence are scheduled.

Source: [Agentic AI Control Planes: The Guardrails That Make Autonomy Operable](/writing/agentic-ai-control-plane/)