Your first agent starts with a conversation
Most owners I meet think an AI agent starts with a tool. Which platform, which plugin, who'll set it up. So they put it off until they have time to learn something technical.
Your first agent starts somewhere much simpler. It starts with a conversation, the same kind you'd have with a new colleague on their first morning.
In our Humane Business Accelerator, I teach it as four steps. Brainstorm, test, then turn it into a skill or put it on a schedule. Here's how each one works, and what to watch for.
Step one. Brainstorm the job
Pick one piece of work that comes round every week and eats more of your time than it should. A weekly update for your team, a summary of new enquiries, the first draft of a client report.
Then talk it through with the AI as you would with a person. Explain what the work is for and who reads it. Say what good looks like and where it usually goes wrong. Ask it what it would need from you to do it well, and answer honestly.
You don't need clever prompts for this. You need to think out loud with something that asks good questions back. Most people find this part useful on its own, because it makes them write down a process that has lived in their head for years.
Step two. Test it until it's right
Now give it a real example and let it try. The first answer won't be right, and that's fine.
The rule I give every group is to never accept the first output. Read it, say exactly what's wrong with it, and send it back. Too long, wrong tone, missed the one number your client cares about. Each round, it gets closer, and you get clearer about what you want.
If you're frustrated at this stage, that's a good sign. It means you know what good looks like and you're still the one driving. The teams that worry me are the ones who stop checking once it starts working.
Keep testing until you'd be happy to send the result with your name on it. For most tasks that takes a handful of rounds.
Step three. Make it a skill
Once the output is right, save what you've learned so you don't have to start again next week. That's all a skill is. As I put it to every group, "The skill's just a prompt."
Write down the brief that worked. Include what the work is for, the steps, the examples of good and bad, and the things it must never do. Save it somewhere your AI tool can reach it. Next week you run the skill and check the result, instead of explaining everything from scratch.
Plenty of work stops here, and that's the right answer. If you only need it now and then, a skill you run when you choose is enough.
Step four. Put it on a schedule, if it earns one
Some work happens on a clock. The Monday team update, the Friday pipeline summary, the end-of-month client report. For these, you can give the skill a heartbeat, a schedule that wakes it up and runs it without you asking.
That's the point where a skill becomes an agent. It has a brain, which is the AI model. It has a memory, which is your files about how you work. It has hands, which are the things it's allowed to do. And it has a heartbeat, which is the schedule.
Don't rush this step. Only schedule work you've tested enough to trust, and keep a person reading the output until you're sure. Scheduling a skill that isn't ready only makes the mistakes arrive on time.
Before any of it, decide what never goes in
There's one job to do before step one. Sit down with your team and agree what you're happy to put into AI and what never goes in. Client files, personal details, anything under a confidentiality agreement.
For anything an agent will run on its own, go further and do the first pass by hand. Map the data it will touch, the decisions it can make alone and the ones where it has to ask you first, and how you'd notice and undo a mistake. It feels slow, and that's the point. You know what an agent will touch before it touches anything.
A worked example
Say you run a small consultancy and every Monday you write a short update for your team. It covers last week's wins, this week's priorities and anything stuck.
You brainstorm it with the AI. You explain who reads it, what they need from it and the tone you use. You paste in two old updates you were proud of.
You test it on last week's notes. The first draft is too formal and misses the stuck items. You say so and send it back. The third draft is close. The fifth is something you'd send.
You save the brief as a skill. For a month you run it every Monday morning and read it before it goes out. When you've stopped making changes, you put it on a schedule for 08:00 on Mondays, and it lands in your inbox to check.
That's an agent. No code, and no clever prompts. Just a conversation you had once, written down and given a heartbeat.
Four mistakes I see
Skipping the brainstorm. People type a one-line request and judge AI on the answer. Give it the context you'd give a new colleague, and the answers change.
Starting with the clever idea. The most impressive agent is rarely the best first one. Start with dull, frequent work where a mistake is easy to spot and cheap to fix.
Scheduling too soon. If you're still changing the brief every week, it isn't ready for a heartbeat.
Keeping it to yourself. A skill only you know about is fragile. Share the brief with your team so they can run it, improve it and notice when it drifts.
Where to start this week
Pick one task. Have the conversation. Test it until it's right, then save it as a skill. Leave the schedule for later.
If you'd like to do this with us, our Humane Business AI Sprint is three hours live on Zoom with Rosario and me. You leave with a vision for the next decade, a roadmap for 2027, your AI set up with the right guardrails and your first agent working. There are three dates in December.
Humane Business is the company I co-founded with Rosario Londoño. Want to build your first agent with us? In the December AI Sprint you build it live, and it's running before you log off. Pick Friday 4, 11 or 18 December.