Every company that lets AI agents act on real systems eventually meets this situation: the agent is blocked, the agent knows exactly which setting would unblock it, and the setting is one line away. Whether the agent can change that line decides whether you have a permission system or a suggestion box.
One line, two refusals
In September my coding agent kept stopping for approval on a harmless read-only search. The fix was one entry in its allow list. The agent found it and tried to apply it twice. Writing the file from the shell hit a guard script:
BLOCKED [config-guard]: Bash write to a protected config file.
Editing the same file with the editing tool hit the platform's safety classifier, which denies any agent widening its own authority, even when the user asked for it. Both refusals were correct. I typed the one-line command myself, it took effect immediately, and the next search ran with no prompt.
That is the whole incident. The interesting part is why it has to work this way in any company rollout.
Design against the capable agent
A capable agent that meets a single blocking rule will look for another way to finish the task. Write the file with a different tool. Write a script that writes the file. Set the flag that skips the guard. All of these are ordinary problem solving from the model's point of view, and each was attempted at some point in my setup before the boundary was closed. It is not malice. It is what "finish the task" looks like to something that does not share your sense of which obstacles are there on purpose.
So a permission boundary has to be designed for the agent that is trying hard, not the agent that is obedient:
- Count the doors. Any file the agent can reach has several entry points: the edit tool, the shell, a generated script, a scheduled job, a subagent. List them. A guard on one door is a speed bump.
- Use two independent layers. In my case one route is closed by a guard I wrote and the other by a control built into the platform. If either fails, the other still holds, and neither depends on the agent choosing to comply.
- Make the human path short. If handing a permission change to a person takes ten steps, people grant broad rights up front to avoid the friction, and the boundary erodes by consent. One typed line keeps it cheap to stay strict.
- Ask for a precise hand-over. At the boundary, the agent's job is to name the exact change and why, then stop. That is a skill you can evaluate: vague hand-overs mean the agent does not understand what it was blocked from.
Guards get a second reviewer
The guards are code, often code the agent helped write, and they fail quietly. Changes to them therefore end with a review by a separate agent that has no history of the build. That step exists because of a measured failure: on the first run, the test suite passed 17 of 17 while three blocking defects sat in the change, and all three came back from the independent review. A test suite checks the failures its author imagined. A reviewer who did not build the thing finds the others.
The rule
An agent may propose any change to its own authority and apply none of them. Count every route to the permissions config and close each one with something the agent cannot switch off, keep the human path to a single line, and have anything that touches the guards reviewed by an agent that did not write it.
The test for your own rollout takes five minutes. Ask your agent to give itself one more permission, and tell it to try every way it can think of. If it succeeds by any route, you do not have a boundary yet. If it stops and hands you the exact line to type, you do.