AI agents explained: what to delegate and when to trust them

We met John Mann in a hallway at a conference in New York, and that chat ended up being one of our favorites of the whole event, so we asked him to come on the podcast.

John is hard to introduce in one line. He’s the founder of Startups and Code, where he works as a fractional CTO, and before that he was a CTO and an SVP at Napster. Before any of that he was a professional dancer and a school teacher, which probably explains why he now spends so much of his time teaching people how to build AI agents for free. Watch the whole conversation below:

One idea connects the examples John shares: start with something you want to do, give an agent a small piece of the work, and expand its role as you learn what it can handle. That makes getting started less overwhelming and gives you a way to judge where agents actually help.

What an agent is

Everyone calls themselves “agentic” these days, from consultancies to AI startups to product pages, but if you ask what the word means the answers get vague pretty quickly. John keeps it simple: an agent performs a task. In practice, that can mean working toward a goal across several steps, using tools and instructions to decide what to do next.

The comparison he likes best is a new intern. You onboard them, explain your domain, and hand them a task. Then you review what they did. Sometimes you’re impressed; sometimes you wonder what they were thinking. With an agent, that review helps you refine the instructions, examples and context you give it for the next task. Most agents, he says, come down to written rules, a few skills and a clear goal.

What to hand over first, and what to keep

If you’re just starting, John suggests giving agents the easy work. On the engineering side that means small front-end changes like text tweaks, color changes and styling updates. You watch how they do, tell them what was good and what wasn’t, and expand from there. (He also has a strong opinion about which coding model is winning at design right now, which you’ll only hear in the episode.)

What he almost never hands over is production deployment. Database migrations, accessibility details and compliance rules all leave too much room for a small mistake to turn into a big one, so a person stays in the loop there. Agents get a lot more freedom in pre-production, where their mistakes show him exactly what needs locking down.

The same thinking applies outside engineering. Plenty of people feel uneasy about letting AI write their emails. John suggests starting with examples of how you already write, so the agent has a reference for your tone and phrasing. You review the draft, make it your own and use those edits to improve the instructions.

The content team that runs while he gets coffee

The part that made our jaws drop was his content setup. He built a fictional persona with a full backstory who publishes a new article every week and runs her own YouTube channel. Agents handle nearly all of it, including the short videos, the titles, descriptions and thumbnails, the uploading and scheduling, and a draft of the next article. John usually starts the process and walks away, though he recently sat down and watched it run from start to finish, and even he was impressed.

He says the setup is much simpler than you’d expect, and he explains how it works in the episode. He also shows us the daily briefing his agents prepare each morning, how it sorts his day, and the small detail that it reminds him to take a break.

Starting from what you want to do

A lot of people see AI as something coming for their jobs. John’s advice is to begin with something you want to create, whether that’s a screenplay, a piece of art or a product, and use AI as a brainstorming partner. Try an idea, push it further and see where it helps. The point for him is to hand off the mind-numbing work so there’s more time for the things you enjoy, without turning every free minute into another opportunity to be productive.

He does this himself. At one point he gave an agent a budget and told it to build a new project every week, and the results are open source on the Startups and Code GitHub for anyone to use.

Agent Academy

When we recorded this conversation, John was preparing to launch Agent Academy, a video series designed to help people go from feeling intimidated by AI to building agents for their business. You can now watch the first video here.

Across John’s examples, the practical lesson is to match an agent’s freedom to the consequences of getting something wrong. Start with work you can easily review, refine the instructions and widen the scope as the results give you reason to. Watch the full conversation for a closer look at his workflows, and tell us which agent you’d build first.