Every "don't ask again" is a delegation. Give it an owner and an end date
Somebody on your team said yes to an agent once, a little tired, and meant it for that afternoon. The tool remembered it as yes forever, for everyone, in places nobody looks. That's how most agent permissions actually get granted.

A standing approval, whether it's a coding agent's allow rule, a chat channel's default repository or a support agent's auto-approve threshold, is a delegation of authority. Treat it like one: name who granted it, what exactly it covers, the limit, when it expires and how it's revoked. Keep the list in one place, review it on a calendar, and let anything nobody claims lapse.
The click that outlives the moment
Most permissions an AI agent holds weren't granted in a meeting. They were granted in a prompt.
A developer is in the middle of a task, Claude Code asks whether it may run a command, and the developer picks "Yes, and don't ask again." Anthropic's documentation is clear about what that does: the rule is saved to a settings file at the root of the repository and applies to future sessions anywhere in that repo. It was a decision about one afternoon. It now governs every afternoon after.
That's not a flaw in one tool. It's what standing approvals are. They trade a little oversight for a lot less friction, which is usually a good trade. The problem is that nobody writes them down as what they are: a delegation of authority from a person to a system.
A permission nobody remembers granting is still a permission. It just has no owner.
Decision guide
A yes for now vs a yes from now on
| Consideration | One-time approval | Standing approval |
|---|---|---|
| Who it binds | This action, this moment | Every future matching action, often for everyone in the project or channel |
| Where it lives | In the person's head | In a settings file, a channel setting or a policy threshold |
| Review | Happens as it's given | Happens only if someone schedules it |
| End | Ends with the action | Lasts until somebody removes it |
Where standing approvals hide
Once you start looking, they're everywhere agents work.
Coding agents. Allow rules for commands, domains the agent may fetch from, tools it may call. Some live in a shared project file, some in a local one, some in a person's global settings. Recent research into Claude Code's local settings file showed how far those rules can travel, passwords included.
Chat-started agents. In GitHub's Slack integration, a channel's default repositories and default model are shared across the channel. Whoever set them decided, for everyone, where work goes when nobody names a repository.
Support and operations agents. A policy that says the agent may ship a replacement under a dollar amount, or refund without review below a threshold. KODIF's own sample policy includes exactly that kind of line: ask for a signed statement above $250. That number is a standing approval. So is whatever the agent may do below it.
Automations. Workflow tools that run on a trigger, with the permissions of whoever turned them on.
None of these are bad. All of them are decisions someone made on behalf of the business, usually without calling them decisions, and often without telling anyone else they'd been made.
A permission nobody remembers granting is still a permission. It just has no owner.
Write it like a delegation
When a manager delegates signing authority, a good company writes down five things. The same five work for an agent.
Who granted it. A name, not "the team." If that person leaves, the approval needs a new owner or it ends.
What it covers. As narrowly as the tool allows. "Run the test suite" is a scope. "Run anything in this folder" is not. For coding agents, that's the difference between approving a pattern and approving a literal command.
The limit. A dollar amount, a number of actions per day, a list of systems it may touch. If the tool can't enforce the limit, say so, and decide whether the approval still makes sense.
When it ends. A date. Most standing approvals don't need to be permanent, and the ones that do should be renewed on purpose.
How it's revoked. Where the setting lives and who can change it. If revoking it takes a support ticket to a vendor, write that down too.
You don't need a new system for this. A short page per team, kept next to the rest of your operating documentation, is enough to start.
One, written out
Here's an illustrative entry for a support agent, the kind of thing that fits in a shared document.
Granted by: the head of customer experience. Covers: replacement shipments for orders marked delivered that the customer says never arrived, US orders only. Limit: order value under $150, one replacement per customer per 90 days. Ends: January 31, then renewed or narrowed after the holiday review. Revoked by: switching off the replacement action in the agent's policy settings; the support operations lead and the head of customer experience can both do it.
And one for a coding agent. Granted by: the engineering lead. Covers: running the project's test and lint scripts. Limit: those named scripts only, no network calls beyond the package registry. Ends: reviewed monthly. Revoked by: editing the shared project settings file, with a pull request.
Neither entry took more than five minutes to write. Both answer the question an auditor, a customer or a new teammate would ask.
Review on a calendar, not after an incident
The weakness of standing approvals is that nothing reminds anyone they exist. So put the reminder in.
In the first week after granting one, the owner looks at what actually ran under it. Approvals usually catch more than intended, and the first week is when that's cheapest to fix. After that, review the whole list monthly. It takes less time than it sounds, because most entries won't have changed. Anything without a current owner is removed. Anything past its end date lapses unless someone renews it with a reason.
Tie the review to events, too. When someone leaves, their approvals get reassigned or removed the same week. When a vendor changes what an agent can do, the approvals that reference it get a second look.
Sequence of work
The life of a standing approval
- Granted
Written down with owner, scope, limit, end date and how to revoke it.
- First week
Owner checks what actually ran under it. Narrow it if it caught more than intended.
- Monthly
Review the list. Anything without a current owner is removed.
- End date
Lapses unless the owner renews it on purpose, with a reason.
Put it into practice
Pick one team this week, probably engineering or support, because that's where agents are already acting. List every standing approval you can find: settings files, channel defaults, policy thresholds, automations. For each, fill in the five lines. You'll find some nobody can explain. Remove those first.
Expect the first pass to turn up surprises. Then make the rule simple enough that people follow it in the moment. "Don't ask again" is fine for patterns that can't do harm. For anything that touches money, customers, production data or credentials, the answer is yes once, and a written approval if it needs to be permanent.
The point isn't to slow people down. It's to make sure that when an agent acts on its own, someone in the business can say who allowed it, and why. We've argued before that some approvals should stay human; this is how you keep track of the ones that don't.
If you're automating real work with agents and want ownership, limits and recovery designed in from the start, that's the core of our business automation practice. Talk to TWIMCO.
Related work: Spin:Market
We build alongside the people doing the work.
Bring us the problem you’re thinking about.