Pillar 03 · Automation
A4 Agent Permissions Standard
An agent should never be able to do more than its owner would sign off.
Form Document and YAML
What it's for
An agent that sends, changes or approves things can make thousands of decisions an hour. If it runs under a person's account or a shared service account, nobody can tell what it did, and it can do anything that account can. This standard sets the permissions every agent must have: its own identity, the least access it needs, limits on what it may do, a log of everything it does, and a tested way to stop it.
What you get
- The standard Seven requirements
- By risk tier More risk, tighter limits
- Permission record One record per agent
- As code A permission manifest
- Checks Before go-live
How it connects
One record per S1 agent: C3's charter is the business agreement it enforces, and S4's check covers the model it calls.
A flavour of the worked example
| Agent identity | svc-agent-invoice-coding |
|---|---|
| Explicitly denied | suppliers: change bank details, payments: any, cost centres: create, change |
| Data class ceiling | Confidential |
| How to stop it | Disable identity svc-agent-invoice-coding in the identity provider |
| Last stop test | Stopped in under 2 minutes; queue held |
A few fields from one record of the toolkit's worked example, with long text shortened.
Next step
Guided build coming next.
Request access
Ask for access to the guided builds, including this kit's when it is ready.