AI Agent Checklist: 20 Things to Check Before Agents Act

This AI agent checklist is the list we would want any team to run before letting an agent change real work. It has 20 checks in five groups: scope, approvals, limits, audit and people. If you cannot tick most of them, keep your agents at "suggest only" until you can.

Key takeaways

• An agent you cannot pause is an agent you should not run.
• Decide what the agent may never do before you decide what it may do.
• Every agent needs a named human owner, not "the team".
• Budgets and action limits go on before the first run, not after the first surprise.
• Start with suggestions, then earn autonomy one level at a time.

How to use this checklist

Paste it into a task with a checkbox per item, and go through it with the agent's owner.

Score it honestly. A "partly" is a "no".

These bands are our suggestion, not a standard. Adjust them to your risk.

Part 1: Scope (checks 1 to 4)

1. The agent has one clear job. "Post a weekly digest" is a job. "Help with the project" is not.
2. You know which projects it can see. It should see the projects it needs, and nothing else.
3. It cannot see more than its owner. If the agent acts for a person, it must not open what that person cannot.
4. The allowed actions are written down. A short list, such as comment, create a subtask, draft a doc. Anything outside the list cannot run.

In AlianHub, each agent is built from a skill, a list of allowed actions and one autonomy level. Nothing outside the list can run, which makes check 4 a property of the tool rather than a promise.

Part 2: Approvals (checks 5 to 9)

5. You chose an autonomy level on purpose. Not the default, a decision.
6. Wider changes wait for a person. Moving work between lists or assigning people should need a yes.
7. There is one place to approve. Proposals scattered across email and chat get missed.
8. Declining gives the agent a reason. A "no" with a reason is how the next proposal gets better.
9. A person closes the work. The agent can move a task to "in review", but a human marks it done.

The approval queue is the product. Everything else is plumbing.

AlianHub's autonomy ladder runs from L0 (answers and suggests) to L3 (acts on low-risk changes and may run on a schedule). Wider changes land in the AI Inbox, where a person approves or declines, and a decline can carry a reason the agent sees. No agent closes a task in any project.

Part 3: Limits (checks 10 to 14)

10. Destructive actions are impossible, not just off. Deleting projects or tasks, removing members, changing permissions and billing should not exist for the agent.
11. There is a daily action limit. A loop should hit a wall, not your whole board.
12. There is a spend cap. Monthly is fine. Set it before the first run.
13. There is a pause switch. One switch that stops every connected agent in the workspace.
14. Sensitive projects can be stricter. A legal or finance project should be able to hold every agent to proposals only.

Tip: ask "what happens if this agent runs 500 times tonight?" If the honest answer is "I do not know", fix checks 11 and 12 first.

In AlianHub, destructive actions are absent from the action registry rather than disabled. Each agent has a daily action limit and a monthly spend cap, a workspace can set a daily AI budget, a project can be stricter than the agent, and one switch pauses every connected agent.

Part 4: Audit and undo (checks 15 to 17)

15. Every action is logged. Acted, proposed and refused, with the agent's name on it.
16. Changes the agent made alone can be undone. One click, not a support ticket.
17. Settings changes are tracked. If someone widened an agent's scope last Tuesday, you can see it.

A log of "what the agent did" is not enough. You also want the attempts that were refused, because those tell you what it tried.

AlianHub writes every agent action to the audit log, whether it acted, proposed or was refused. Changes an agent makes on its own come with Undo, and each agent keeps a revision history of its settings.

Part 5: People (checks 18 to 20)

18. The agent has a named owner. One person who reviews its work and answers for it.
19. The team knows it exists. Nobody should discover an agent by reading its comments.
20. You have a review date. Two weeks in, look at what it proposed, what was approved and what it cost. Then decide to promote, keep or retire it.

If nobody owns the agent, everybody will blame it.

How to do this in AlianHub

Most of the checklist is a setting you can tick in minutes:

1. Choose the agent. In AI, open the Agents tab, or turn on a team pack. Pack agents start paused.
2. Check scope and skill. Confirm its skill, its allowed actions and its project scope (checks 1 to 4).
3. Set the level. Pick L0 or L1 for the first weeks (check 5).
4. Set runs and limits. Enter a daily action limit and a monthly spend cap (checks 11 and 12).
5. Dry run the skill. In the Skill library, use Dry run to preview what it would do. Nothing is written.
6. Make sensitive projects stricter than the agent, so they stay proposal-only (check 14).
7. Review in the AI Inbox and read the audit log after the first runs (checks 7, 8 and 15).

A business scenario

An operations team wants the Stock Alert and Delivery Tracker agents from the Operations pack. Before switching them on, the lead runs the 20 checks: one owner, one project, L1, a spend cap and a review in two weeks.

At the review, the AI Inbox shows which proposals were approved and the audit log shows nothing outside scope. The team promotes one agent to L2 with confidence instead of arguing about risk, and late deliveries reach a person sooner.

We built AlianHub so that most of this checklist is the default rather than homework. You can see the autonomy ladder and skill library on the AI page, and the hard stops, pause switch and health views on the AI controls page.

If your team prefers to use Claude or ChatGPT, the same idea applies: connected AI acts as you, sees only what you can open, and cannot delete tasks, docs, comments or projects.

For the bigger picture behind this list, read our view on the future of AI in project management.

FAQ

What is an AI agent checklist?

It is a list of checks to run before an AI agent is allowed to change work in your tools. It covers what the agent can see, what it can do alone, what needs approval, and how you stop and undo it.

Which check matters most?

Check 10. If destructive actions are only switched off, someone can switch them on. If they do not exist for the agent, they cannot be reached.

How long should an agent stay in "suggest only" mode?

Long enough to judge its proposals. A week or two of reading its suggestions is a reasonable start before allowing low-risk changes.

Can I test agents without risk?

In AlianHub, skills have a Dry run that previews what they would do without writing anything. You can also explore the live demo.

Run this checklist with your team this week, then try the agents in the live demo or self-host AlianHub for free. If it helped, share it with whoever is about to switch on their first agent.