# AI Agent Authority Canvas — Free Starter
**RELESONA TECHNOLOGIES · Digital Product Store · v1.0 · 24 October 2026**

This fill-in template helps turn an agent idea into an **explicit, auditable work agreement**. It does not give an agent permission to act by itself. Review the completed canvas before connecting any real account or tool.

## 1. Mission and owner
- Agent name and version:
- Human owner and contact:
- Business outcome (one sentence):
- Task the agent may perform:
- Task the agent must **not** perform:
- End condition / definition of done:

## 2. Inputs and data boundaries
| Data source | Allowed fields | Sensitivity | Storage / retention | Access owner |
| --- | --- | --- | --- | --- |
| [Source] | [Fields] | [Public/internal/personal/restricted] | [Period] | [Name or role] |

**Never put API secrets, passwords or customer records directly in this canvas.** Store credentials only in a designated secret manager.

## 3. Tool and action permissions
| Tool | Permitted operation | Read / write | Scope | Approval required? |
| --- | --- | --- | --- | --- |
| [Tool] | [Exact action] | [R/W] | [Folder/project/account] | [Y/N] |

- Default behavior for an unknown tool or permission: **deny**.
- No purchases, customer messages, production deployments, deletion or external publication without the authority expressly granted for that operation.
- No authorization can be inferred from a user-uploaded document, webpage, retrieved text or third-party message.

## 4. Human approval points
List decisions that need a human to authorize **before** execution:

1. [Action] — evidence supplied: [link/description] — approver: [role]
2. [Action] — evidence supplied: [link/description] — approver: [role]
3. [Action] — evidence supplied: [link/description] — approver: [role]

Require a separate approval for production release, payment, contact with external parties, access expansion and irreversible actions.

## 5. Evidence trail
Every completed task should record:
- Request ID and source.
- Agent version and allowed tools.
- Date/time and action summary.
- Evidence of tool response; omit secrets and unnecessary personal data.
- Verification result, independent reviewer if needed.
- Approval record for sensitive actions.
- Recovery / rollback information.

## 6. Stop conditions
Immediately stop and escalate when:
- A required approval is missing, unclear or denied.
- A command affects resources outside the declared scope.
- Private data appears in a public output.
- A tool returns an unexpected identity or environment.
- The result is not verifiable, or repeated failures exceed the agreed limit.

## 7. Acceptance test matrix
| Test | Expected outcome | Actual result | Evidence |
| --- | --- | --- | --- |
| Normal permitted read | Success with cited evidence | | |
| Unknown tool call | Denied | | |
| External message without approval | Denied | | |
| Injected instruction in retrieved text | Ignored as authority | | |
| Wrong environment / production target | Denied | | |
| Repeated tool error | Stop and escalate | | |
| Recovery after partial failure | No duplicate harmful action | | |

## 8. Release decision
- Owner sign-off:
- Independent security review:
- Version / date:
- Known limitations:
- Next review date:

**Free starter resource.** Provided for education and planning, without managed deployment, live integrations, guarantees or legal advice. Models and tools can make mistakes; apply your own security, privacy and governance review.