The agent onboarding plan
This is the companion guide to the video about what to do when you are told to supervise AI agents. It is a one-page plan you fill in: what the agent can touch on day one, how closely to check it in week one, how to step back one task at a time, and the short list that never moves. If you have not watched yet, start there.
Episode 6Watch on YouTubeThe one idea
“Human in the loop” sounds like a policy, but it is a phrase. A real plan answers four questions. What can the agent touch? How closely do I check? How often? And what never moves up? You already answer those when someone joins your team. Answer them for an agent too, in three stages: scope on day one, checks in week one, then step back one task type at a time, only where it has earned it.
The one-page plan
Copy this, fill in the brackets, and keep it next to the agent’s work. Do it before the agent touches anything real. For a person, this is the onboarding doc. For an agent, it is the same doc.
AGENT ONBOARDING PLAN Agent: [name of the agent] Owner (the person who answers for it): [name] Start date: [date] Its one job, in one sentence: [job] How I will know it is doing well (a number): [target] STAGE 1. DAY ONE: SCOPE It can see: [accounts, folders, tools] It can change: [what it may edit, create, or send] It can never do alone: [the short list] First task (read only, or easy to undo): [task] Rule for anything that goes outside: it shows me the exact recipient and the final wording first, unless I have given standing permission for that exact kind of message. Where it reports to me: [channel, time of day] STAGE 2. WEEK ONE: CHECK EVERYTHING I check every output. Daily check time: [time] Trial run first: the first time it does anything outbound, it does it to me, with made-up details. Three questions for every output: 1. Is it right? 2. Is it complete? 3. When it was not sure, did it tell me? Log (two minutes a day): date | task | what was off | what it changed STAGE 3. STEP BACK, ONE TASK TYPE AT A TIME Task type: [one kind of task] Clean checks in a row before I check less: [number] Then I check: [a sample, once a week] Way back: if it slips, that task type goes back one step. Not everything. WHAT STAYS WITH ME - Anything it submits as me - Any new list or new recipient, unless I have approved that exact kind - Anything with money - Anything I have not reviewed at least once - Instructions it finds on web pages. It follows what I tell it, not what a web page tells it.
What it looked like in practice
I ran this plan with an AI agent over about five weeks of real work. The dates matter more than the details.
- Day 1. I asked for the access list before giving it a task. The first task was read only. I approved seven of its ten suggested inbox rules, and it said it would ask before sending anything to anyone else.
- Day 5. The first real outreach started with a trial email to myself. A bug then put a blank greeting in a batch of twenty. It paused the rest, explained what broke, and we added a new rule: one test email first, and stop on the first mismatch.
- Day 8. A standing assignment: one research project a day.
- Day 13. One task type got standing permission, because the evidence said it was safe.
- Week 4. New batches went out on schedule without asking me first. The short list at the end of the plan stayed with me.
The question to ask after a mistake
In week one, the most useful habit is the one you would use with a new hire. When something is wrong, ask what caused it and what changes.
AFTER A MISTAKE
Do not judge the agent on whether it apologizes. Ask this:
"What caused it, and what do you change so it does not happen again?"
A specific answer ("I only read the first page of the staff directory, so I will read the whole directory before I call anything blank") is a good sign. A vague one is the real red flag.When you take work back
Stepping back is not the only direction. Sometimes you take a task back, and it is usually not because the agent failed. There are three common reasons:
- Blocked. A site refuses an automated visitor, or a captcha stops it.
- Throttled. A platform limits how much an agent can do, so a person finishes the last step.
- Faster yourself. You are already logged in and it takes you ten seconds.
Managing an agent includes deciding which tasks are its job and which are still yours. For the checklist I used before giving an agent access, see the agent checklist from Episode 4. For setting up an agent that signs in as you, see your first week with Muse.
One caution if you work at a company
Check your company’s AI policy before you connect work email or shared drives to any agent. Day one is also when you find out what you are allowed to connect.
Want an onboarding plan built for your team’s agents?
We design and deploy AI agents with the approvals, the access limits, and the step-back rules built in from the first day.
Start a Conversation