How to build an AI agent team: recap of the deep dive

Most people who added AI to their week ended up with more work, not less. The reports say it, and the rooms I teach in confirm it. The reason is almost always the same: the tool is being used like a chatbot when it should be used like a colleague. You ask, it answers, you copy the answer somewhere, you ask again. That is a faster typewriter, and you are still the one doing every piece of the work.

This is the recap of the AI agent deep dive I ran for the WorkSmart community. The goal of that session, and of this post, is one shift: stop babysitting the tool. You should be able to close the laptop, go to sleep, and wake up to work that already happened.

What is an AI agent, in plain terms?

An agent is software that can perform tasks on its own, make decisions inside the boundaries you set, and take actions on your behalf.

The difference shows up in how you talk to it. The chatbot version sounds like this: read this lease, tell me what to say, draft the email. The agent version sounds like this: I want out of this lease, make it happen. Then it starts working the goal, and it comes back when the goal is met or when it hits a wall.

That is the mental move. Stop assigning tasks. Start assigning goals. If you want the longer version of the definition, what AI agents really are covers it from zero.

The four levels of autonomy

Agents are not on or off. Think about it the way the car industry thinks about self-driving, in levels, and pick your level deliberately for each job.

I have seen entire companies run at the fourth level, staffed by agents that emulate people convincingly. I do not recommend building that. I do want you to know how it works, because the level you choose should be a decision you made, not a default you inherited.

If you have read what AI agents really are, you have seen me describe a ladder that goes chat, then workflow, then agent. That ladder is about how much the software can do. These four levels are about how much you let it do without asking. Same territory, two different questions, and you need an answer to both.

The agentic loop, and why "done" is the whole game

Every agent runs the same shape:

1. Observe. Something happened, or the schedule fired, or a piece of work came back.
2. Plan. It writes out how it intends to handle this. In most tools you can read that plan.
3. Act. It does the work.
4. Check. Did that satisfy what she asked for? If no, back to step two.

The loop is the reason I keep hammering on one word: done. A loop closes when the finish line is documented. If you have not defined what finished looks like, the agent will keep working, keep burning tokens, and keep spending your money on a job it cannot recognize as complete.

I think about it the way I think about my human team. If I tell them at the top of the year to go make a million dollars, that loop does not close until the million is in. Nobody comes back to me at $600K asking if that counts. The end state was defined. Give your agents the same clarity, and read token efficiency if the cost side of open-ended loops is new to you.

What this looks like when it is running

Here is my actual setup, so you have something concrete to picture.

At 6:34 in the morning, before I am awake, a Slack message lands. It tells me what is on my calendar, what moved in my inbox, and what needs a decision from me. Read it and it sounds less like an email digest and more like a status report from a small organization that worked overnight.

• One agent owns my personal revenue: brand deals, speaking, WorkSmart. It reports whether anything moved.
• Another acts as a digital revenue officer on the company side, checking whether any client came back in the last 24 hours and whether anything in the pipeline I personally work moved. When a client does come back, it updates a spreadsheet row and drafts the response.
• Another reads my inbox and pre-sorts it with labels: deals, follow-ups it has already drafted, misrouted mail that belongs to a different business of mine, things that genuinely need me, PR pitches, and one folder called uncertain for anything it could not classify.
• Another reads the notes from meetings I was not in. Our sales team meets weekly. I am not in that room, but I know every deal status, and my agent flags the gap: two weeks ago someone said the Treehouse deal was about to close, and today they did not mention it at all. Ask about that.

None of that requires me to open a laptop. The agents run in the cloud on a schedule. The morning brief is the output, and the folders are the filing. My part is reading and deciding, which is the part that needs me.

That last example is the one people underrate. Meeting transcripts plus an agent that holds three or four meetings in memory is one of the highest-value setups available right now, and it is simple to build. Using AI for meeting notes is the on-ramp.

Cloud or local? Pick per job

This is new and it matters. Scheduled agent work can now run in the cloud, on your own space, whether or not your computer is on. That removes the old annoyance where your scheduled tasks all fired the moment you opened your laptop.

Keep a job local when it needs files that live on your machine. If your weekly revenue report is built from a folder of downloads, a cloud agent cannot reach them, so that one stays local. Keep it local for privacy reasons too, when the data should not leave your machine at all.

Everything else can move up. Slides in Google Drive, email, calendar, your CRM: all of that is already in the cloud, so the agent can work it while you are on a plane.

Deciding what to hand over

The methodology here predates AI. It is the WorkSmart split between CEO tasks and operating tasks.

CEO tasks are the ones only you can do. For me, that includes sending my own email. An agent drafts it, and I press go. I sign my own contracts. I do not have automation reviewing or signing legal documents.

Operating tasks are everything a competent other person could do while you stayed comfortable. Updating your LinkedIn. Finding leads and sending the first note. Keeping your kids' calendars straight. Building the quarterly supply order from the list your school sends you. If money were no object, what would you hand off? That list is your candidate pool. What to delegate first walks the exercise properly.

Then run each candidate through three tests:

Email passes all three. Calendar passes. Brand deal tracking passes. Keeping tabs on my team's pipeline passes. That is why those got built first.

These three tests are about whether a job can be automated well. There is a second filter about whether it should be, which sorts by how much a mistake costs. The first tasks to give an AI agent runs that one. Use both.

Onboard an agent the way you onboard a person

Give it the same context you would give a new hire on day one. Who are they, what do they do, who do they serve, what problem are they solving, how do they communicate, and how often.

Then show, do not just tell. If you are building a travel agent, do not describe a good itinerary. Hand it a finished one. Say: check the weather where I am going, check me in for flights, alert me to delays, and tell me if the invite says the dress code is something I need to pack for. Here is what the output should look like when you are done.

Give it the tools it needs to find that information. Then let it run a draft on the last two trips, come back, and talk it through with you so you can tune it in real time.

And say what it must never do. Never change a date on my calendar. Never rebook a flight. These tools can take over a browser and click through a checkout, so the boundaries need to be explicit, not implied. AI agent guardrails has the full list of what should stay behind a human approval step.

The team structure: router, specialists, reviewer, memory

One agent is useful. A team is a different level, and the structure is what makes it work.

The router, or chief of staff. Everything goes to this one first. It knows you best, it uses the strongest model, and its job is to break a goal into pieces and hand them to the right specialist. Mine is named Malcolm.

The specialists. Narrow jobs, well trained, running on cheaper models. My graphic designer agent knows the WorkSmart dimensions, colors, logos and fonts. My Substack writer knows my voice. Neither of them needs to be brilliant. They need to be specific.

The reviewer. Right now, you are the reviewer. You do not have to be. A reviewer agent takes the finished work and checks it against the definition of done the router handed over. If it fails, it goes back to the specialist and around the loop again until it passes.

The memory keeper. Mine is named Thurgood, and his only job is documentation. When something changes, when I give a new rule or a piece of feedback, he writes it down so the lesson survives. Without this role, you teach the same correction forever.

Here is that team working on a real job. I take a walk, talk out an idea for a Substack post about which tools I am using and which I dropped, and hand the transcript to Malcolm. He routes the draft to my Substack writer, sends the graphics to my designer agent, and drafts a Slack message to Martin, my actual human assistant, asking for the screenshots I want in the post. Done, to me, means the piece is sitting in Substack as a draft with the graphics in place. That is what came back.

If the org chart framing is useful, the org chart of one maps it out for a solo business.

Build the email agent, step by step

This is the one to build first, so here is the whole thing. Do not skim the prompt and rewrite it from scratch. The length is the point. A short prompt gets you a filter that sorts by keyword. A brief like this gets you something that reasons about what a message means before it decides what to do with it.

Step 1. Connect the tools. In the desktop app, connect your email, your calendar, and one place you want to be told things. Gmail or Outlook both work. For the telling, Slack is what I use because I live there. Email to yourself works fine. If you want it on your phone, an agent can text you through iCloud or a Google Voice number you set up for this.

Step 2. Set the level. Semi automated. It reads, decides, labels, and drafts. It does not send. Turn off any tool that lets it send mail before you run it once.

Step 3. Give it the brief. Paste this, and swap the bracketed parts for your real business. Every bracket you leave generic costs you accuracy on the judgment calls.

You are my inbox chief of staff. Your name is [pick a name] and you work for me every morning before I wake up.

MY CONTEXT
I run [what your business is and how it makes money]. My revenue comes from [list your actual revenue sources, most important first]. My clients and prospects are [who they are]. The people whose mail always matters to me are [names, roles, or domains]. I currently have [rough inbox volume] a day.

YOUR GOAL
When I open my laptop, my inbox should already be sorted, the things that need me should be obvious, and the replies I am going to send should already be written in my voice waiting in drafts. My morning should go to decisions. The reading is your job.

HOW TO THINK, NOT JUST SORT
Do not classify by keywords or sender alone. For each message, reason about two things before you decide anything:
1. What does this person want from me, including what they want but did not say directly?
2. What happens if I never respond? Who is affected, what does it cost, and how fast does the cost grow?
Sort by that second answer. A polite message from a client whose contract renews in three weeks matters more than an urgent-sounding message from a stranger. Judge consequence, not tone. If someone is being vague because they are uncomfortable asking for something, name that in your note to me.

REVENUE IS THE PRIORITY
Nothing tied to money is allowed to slip. That includes new inbound interest, existing clients, renewals, invoices, late payments, partnership or sponsorship feelers, and anything that reads like a warm introduction. When you are not sure whether something touches revenue, treat it as if it does and tell me why you were unsure.

WHAT TO DO EACH MORNING, IN ORDER
1. Read everything that arrived in the last 24 hours.
2. Label each thread. Use these labels: Needs me, Revenue, Drafted, Scheduling, Noise, Misrouted, Uncertain. Create them if they do not exist.
3. Draft a reply for every thread where you can tell what my answer would be. Leave them in drafts. Never send.
4. Cross-check my calendar. If a thread refers to a meeting, a deadline, or a date I have committed to, say so in the summary.
5. Send me one message in [Slack or email] at [time]. Not a list of my email. A briefing.

WHAT THE BRIEFING LOOKS LIKE
Lead with the three things that need a decision from me today, each in one sentence, with what you recommend and why. Then anything revenue related that moved, and what you drafted for it. Then what you handled and filed with no action needed, counted, not listed. Then, last, anything you could not figure out, with your best guess and the specific question you need answered. Keep the whole thing short enough to read before coffee.

HOW TO WRITE IN MY VOICE
Study my sent mail before you draft anything. Match the real length of my replies, how I open and close them, and how direct I am with different people. I write short. I do not use em-dashes. I do not pad with pleasantries. When you draft, write what I would write, not what a polite assistant would write. If a thread needs a no, write the no.

WHAT YOU NEVER DO
Never send an email, reply, or calendar invite. Never accept, decline, move, or delete anything on my calendar. Never delete or archive a message, including things you labeled Noise. Never unsubscribe me from anything. Never commit me to a date, a price, or a deliverable, even in a draft, unless I have already agreed to it somewhere you can point to. If a thread involves money leaving my accounts, legal language, or anything I would sign, do not draft it. Flag it and stop.

WHEN TO WAKE ME UP EARLY
If something is genuinely time-critical and waiting until morning would cost me the opportunity, message me immediately instead of holding it for the briefing. Use your judgment about what qualifies. Getting this wrong twice in a week means you are too twitchy, so tell me if you are unsure where the line is and we will set it together.

DEFINITION OF DONE
You are finished when every thread from the last 24 hours has a label, every actionable thread has a draft or a stated reason it does not, my briefing has been sent, and the Uncertain folder has your best guess attached to each item. If you cannot finish because a tool failed or access broke, say exactly what broke in the briefing instead of skipping it without telling me. Do not keep working past done.

HOW YOU GET BETTER
Keep a running file of what you learn about my preferences. Every time I rewrite one of your drafts, edit a label, or tell you that you misjudged something, write down the rule you should have followed and apply it from then on. Show me what you added to that file at the end of each week. I should have to correct you about a given thing once.

BEFORE YOU START
Ask me every question you need answered to do this well. Do not guess at anything important. Then show me your plan, run it against yesterday's mail only, and let me react before you go on a schedule.

Step 4. Answer its questions honestly. It will ask things you have not thought about, like which of two clients wins when both need you. Those answers are the actual value. Do not rush them.

Step 5. Run it on yesterday only. Read every label and every draft. You are looking for judgment errors, not typos. If it flagged a newsletter as needing you, or filed a real prospect as noise, tell it why it was wrong in terms of consequence, not in terms of keywords. Two or three rounds of this is normal.

Step 6. Schedule it in the cloud. Pick your time. Give it a full week before you touch it, then read a week's worth of drafts against what you sent yourself and hand it the pattern you notice.

One more thing, and it is the part people skip. Two weeks in, ask it this: based on what you have seen in my inbox, what am I spending time on that I should not be, and what would you take over next? It has read more of your mail than you have. Its answer is usually right.

Two habits that will save you money and pain

Keep the job descriptions in files you own. Mine live on my computer, not inside any one platform. That means the same agent works in whichever tool I am using that week, and it means I am not loyal to a company. I am loyal to getting the work done. Platforms now let you import projects, skills and instructions from each other, and hosting your master files in a folder you control makes those moves boring instead of painful.

The rule that goes with it: whenever you give an agent a new instruction, part of the loop has to be updating that master file. Otherwise you end up with the same agent running two different sets of rules in two different places, and neither one is right.

Match the model to the job. Do not put your most expensive model on a task a mid-tier model handles fine. Once, it does not matter. A million times, it does. This is exactly why the router should be the smart one: it decides which specialist, on which model, handles what. The savings compound, for you and for anyone you are billing a flat fee.

Where I draw my own lines

I do not connect my bank accounts, my brokerage, or my health data to any of these tools. I will talk through a scenario in the abstract. I will not hand over integrated access. None of these companies has demonstrated a moral compass that earns that, whatever the marketing says.

Two practical steps in the same spirit. Pay for the business or enterprise tier, which gives you more protection over what can be trained on. Then go turn off marketing personalization and marketing measurement in your data controls, because it may well be switched on right now even on a paid plan. That setting is how sensitive context becomes an ad. AI privacy basics covers the rest of the sweep.

None of this makes AI a bad tool. It makes it a tool, one of a hundred you use to grow a business. Small business owners, and particularly women and Black owners who are chronically underfunded and under-resourced, need every lever that levels the field. Use the lever. Read the documentation. Keep humans in the loop where the cost of being wrong is high. Human-centered AI is the longer argument.

Hand over the work. Keep the decisions. Write down what done looks like, or the loop never closes.

Do this next

Pick one job that repeats every week, that you can describe well, and that has a clear finish line. Email triage and calendar prep are the two best starters. Build that one agent, schedule it in the cloud, and let it run for a week before you add a second. Then build your router, so the second and third agents have someone to report to besides you. The monthly live AI trainings inside the WorkSmart community walk through these builds with real businesses in the room, and every replay and template is in there with them.