Approve an agent in one meeting by walking the checklist in seven sections and recording an answer, an owner and a date for each gap
Facts read 10 Nov 2026. Page reviewed 16 Nov 2026.
A production review for an AI agent works when it follows a fixed order and ends with recorded decisions. This guide gives the pre-read, the agenda built on the 30-check Agent Boundary Checklist, how to score a check, what to do with none found, and what the meeting must leave behind.
The pre-read is a one-page description of the agent
Ask the builder for a page before the meeting, not a slide deck. It names the platform (the map has the entry for each), the identity the agent runs as, the roles attached to it, every tool with the identity each tool runs as, the data stores it touches, the network it sits on, and the person who owns it. Without the page, the meeting spends its hour discovering facts that the builder could have written down in twenty minutes. With it, each of the 30 checks on the checklist becomes a yes or no question that someone in the room can answer from the cloud console, not from memory.
Four people need to be in the room
Invite the person who approves, the person who built the agent, the owner of the cloud platform it runs on, and the owner of the data it touches. The approver decides. The builder answers questions from the pre-read and shows settings. The platform owner can show the account-level controls (ceilings, logging, network) that the builder cannot see. The data owner says whether the access described is acceptable for that data. If one of the four is absent, mark the checks that need them as open instead of guessing. A review held without the platform owner is the usual reason the ceiling and logging checks are marked met without anyone looking.
Walk the checklist in its seven sections, in order
The checklist groups its 30 checks into identity, permissions, data, network, tools, change control and logging. Take them in that order, because the early sections decide how much weight the later ones carry. If the agent runs as a shared default account, the permission answers are about every workload on that account. Start with one identity per agent. Move to the permission policy and the ceiling above it. Then cover data, network, tool authority, testing before enforcing and attributing logged actions to the agent. Each check has a how-to for AWS, Google Cloud and Azure with a provider source.
Allow roughly an hour for a typical agent. A table helps the chair keep time. The counts below come from the checklist. Move on when the time for a section is used, and mark whatever remains as open instead of running over; the open list is the output of the meeting, and a long open list is a finding in itself, and it should reach the approver in writing before anyone signs.
| Section | Checks | Share of the hour |
|---|---|---|
| Identity | 5 | 12 minutes |
| Permissions | 5 | 12 minutes |
| Data | 4 | 8 minutes |
| Network | 4 | 8 minutes |
| Tools | 4 | 8 minutes |
| Change control | 4 | 6 minutes |
| Logging | 4 | 6 minutes |
Score each check as met, not met or open
Use three answers. Met means someone showed the setting in the account during the meeting. Not met means the setting is absent or wrong. Open means the documentation gives no way to meet the check, or nobody in the room can show it. The checklist itself contains three open cells, each reading none found on 2026-11-10: limiting who can pass or impersonate the role on Azure, avoiding default accounts on Azure, and tying the identity's removal to the agent on AWS. Treat those as open and assign an owner to find out. A provider's silence in the pages we read is not proof that the control is missing, and it is not proof that it exists.
Hold a few facts the providers document about defaults
Three documented defaults tend to settle arguments in the room. AWS says to use broad deployment permissions only for development and testing, and that BedrockAgentCoreFullAccess grants broad permissions. Microsoft documents that a new Foundry agent identity starts without resource permissions and that Foundry does not assign RBAC roles to it, so a pass on a permission check means someone granted roles on purpose. AWS documents that CloudTrail trails do not log data events by default, and Google documents that Data Access audit logs are disabled by default for most services, so a claim that everything is logged needs the setting shown (check).
Leave the room with owners, dates and a trim review
A review that ends with "approved" and nothing else cannot be revisited. Record four things. The decision: approved, approved with conditions, or not approved. The conditions, each with an owner and a date. The checks scored open and who will close them. A date for the first trim of unused permissions (check), set after the agent has run on real traffic. Store the record next to the pre-read and the dated provider pages used as evidence. When the agent is later retired, the same record tells the platform team which identity and access to remove (check).
Related pages
Sources
- AWS: IAM permissions for AgentCore Runtime, read 10 Nov 2026
- Microsoft: agent identity in Foundry Agent Service, read 10 Nov 2026
- AWS: logging data events in CloudTrail, read 10 Nov 2026
- Google Cloud: enable Data Access audit logs, read 10 Nov 2026
- Google Cloud: agent identity on Agent Runtime, read 10 Nov 2026
Last reviewed 16 Nov 2026.