Part 31: A cross-vendor agent governance checklist
Part 31 · One Agent, Five Platforms Part 30 mapped one policy model onto five vendors. This part turns it into something you can use in a review meeting: ten questions to answer before any agent goes live, on any platfo
Part 31 · One Agent, Five Platforms
Part 30 mapped one policy model onto five vendors. This part turns it into something you can use in a review meeting: ten questions to answer before any agent goes live, on any platform.
There are no new platform facts here. Every question comes from an earlier part of the series, and each one says where. If the answer to any of them is "not yet", the agent isn't ready.
The ten questions
1. Identity: which identity does the agent use?
A good answer names one identity that belongs to this agent only. It isn't a shared service account, and it isn't whoever is logged in. When the agent acts on behalf of a user, the logs show both. Part 23 compared how Azure and AWS do this, and Part 30 showed which platforms need you to build it yourself.
2. Tool list: which tools can it call, and with what scope?
A good answer is a written list, enforced outside the prompt. For the invoice agent, that's the six reads and three writes from Part 7, switched on stage by stage as in Part 9. Anything not on the list is blocked at the gateway or the process API.
3. Approval gates: which actions need a person, and where is that checked?
A good answer is: every write needs an approval_id, and the process API rejects a write without a valid one (Part 6). The prompt can ask for approval, but the API is what enforces it.
4. PII: what personal or sensitive data can it see, and where can that data go?
A good answer lists the data classes each tool returns, such as supplier bank details or contact names. It also says which filters apply to model traffic and which regions the data may reach.
5. Token budget: what's the limit, and what happens when it's hit?
A good answer is a limit per agent, set at the gateway from the start, as in Part 11. Hitting it parks or fails the request with a named error. It doesn't silently cost more.
6. Audit: can you reconstruct any decision a month later?
A good answer is a case id on every call and message, and a stored record of what the agent proposed, who approved it and when. That record has to survive platform boundaries, including the A2A calls from Part 28.
7. Owner: who is accountable, and who gets the alert at 2am?
A good answer names a team, not a person who might move on. The owner is listed in the registry of record, next to the agent card from Part 29.
8. Kill switch: how fast can you stop it?
A good answer has two levels. One drops a single exception type back to shadow, as Part 9 described: configuration, not code, in minutes. The other stops the whole agent and has been tested. Part 39 covers the second in detail.
9. Eval threshold: what result must it reach, and when is it re-run?
A good answer uses the golden set from Part 10. The exact checks must all pass, including the embedded instruction in GS30. It's re-run after every change to the model, prompt, tools or platform, and a failed run moves that exception type back to shadow.
10. Rollback: can you go back to the last good version?
A good answer covers the prompt, the tool list and the agent configuration together, not just the code. Some platforms document this and some don't: Azure's Foundry snapshots versions with rollback (Part 22), while Boomi's versioned packages had no rollback procedure documented in what I read (Part 15). Where the platform doesn't provide it, keep the configuration in source control and redeploy.
How to use it
- Ask all ten before go-live, and again when an agent moves up a stage or changes platform.
- Write the answers down in the registry of record, so the next review starts from them.
- Treat "the platform handles it" as a question, not an answer. Part 30 showed that the same control can be built in, partial or missing, depending on the vendor.
My guess is that the questions people skip most are the dull ones: the owner and the rollback. They only matter on a bad day, and then they matter more than anything else.
Which question gets skipped most?
This series describes a reference build on a fictional company. This part is a checklist based on Parts 6 to 30, which are documentation-only; no hands-on build has been run yet. Views are my own.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.