Agent teams 101: How to build a small army of AI helpers

“I never would have thought in a million years I’d be able to put up a website and keep it current, or that I’d ever have time to write regularly on pretty complex topics and experiences,” says Kimberly, who doesn’t consider herself a ​“technology wizard”—she’s just not afraid to ask AI a lot of questions to figure out how to use it to support her. ​“I have a fully functioning team running every day. All of that came through working with the tools.”

While Kimberly’s use case is specific, the potential applications of this workflow are vast. AI agents can help you tackle ideas that felt too daunting before, taking on anything especially time-consuming or administrative to give you time to focus on more creative or analytical work. This kind of support could help you ship the microsite you’ve been wanting to launch to help market your brand. The newsletter that you think would be wildly useful for growth but would never have the time to run. Even just the personal dashboard or competitive research briefing that will help you do your job better.

“There are so many things that you would never think you’d have the capacity or time to do that you just can,” says Kimberly.

Here are the exact steps and copy-paste prompts anyone can use to start building up a team of agents to tackle your biggest backburner projects.

What is an AI agent?

To a beginner, building agents can sound like AI on hard mode, but Kimberly describes them simply: ​“An agent is just a written set of instructions telling Claude who to be for one specific job.” (Kimberly built this using Claude, but you can build a similar agentic structure using other LLMs.)

Unlike automation, where you set up the exact steps for a job to get done, an agent starts with just a goal and a role, and it works out the steps within the boundaries you set.

The magic, Kimberly says, isn’t in a single agent, but in ​“splitting your business into narrow roles so each conversation has one clear job, instead of one giant chat that tries to do everything and forgets what matters.”

How to build a team of AI agents

Step 1: Decide on your team structure

Like building any team, the first step is figuring out what roles you need agents to play. Generally, Kimberly says, you’re aiming for a ​“hub and spokes” structure: One coordinator (like a chief of staff) which you actually talk to day-to-day and which manages all the other agents, and then specialists which each own a narrow function of the execution.

For the specialists, Kimberly says you should be able to describe the outcome they own in one sentence. ​“If you need two sentences to describe the job, you probably need two agents,” she explains.

For instance, on her ​“team,” one agent is in charge of searching for relevant content on the web, another decides which of that content is most interesting and relevant to her work, and a third writes summaries for that content and flags recommendations for topics that could make interesting newsletters or articles.

Kimberly says there are two ways to go about determining the staff you need. The first is what she did: Talk to your chatbot of choice about what you are trying to do and ask what teammates you need.

“I asked Claude, ​‘What would you recommend as the right org structure to manage all of the work that I’ve been doing for this website?’” she explains.

You can also start by building your coordinator and one specialist for the task you most want to offload, and then build others as you see the need.

Step 2: Write a job description for each agent

Writing instructions for an AI agent is similar to writing a job description: You want to give info on what they own, what they don’t, basics about your business or project, plus any instructions for how they should work or the standard outputs you expect from them. Kimberly says this part is actually quite easy: if you’re clear on the roles, they almost write themselves.

For example, here are some sample instructions for a coordinator agent:

And some for a specialist pipeline agent:

Step 3: Tell your agents when to keep a human involved

The thing that’s different about working with agents than a normal chatbot is that, theoretically, they can do work without you ever looking at it. 

“For most of the time I have spent with these tools, I ask something, the tool answers, and I decide what to do with the answer,” explains Kimberly. ​“The judgment stays with me, because the tool has not actually done anything, it has only handed me material to work with. Building a staff of agents is a different posture entirely. The point is to hand them standing work and let them carry it out without me watching each step.”

Kimberly wanted to make sure her agents never took that autonomy too far, so she set up what she calls ​“escalation barriers” as part of every job description—things the agent must never do without stopping to check with her first. 

Coming up with the language for these escalation barriers also involved a conversation with Claude where Kimberly walked through all of her worst-case scenarios and asked what language needed to be in the skills document to prevent that.

Here’s Kimberly’s template for a full agent job description with the escalation barriers built in:

Step 4: Make sure your agents can work with each other, and with the tools they need to do their jobs

The simplest version of this build can be done in the free Claude plan using Projects, and Kimberly explains how to do so as part of her guide here. The downside is, with that version, the agents can’t actually coordinate with each other, so you’ll need to be the go-between.

The upgraded version builds your agents as Skills in Claude’s paid plan, allowing you to only interface with the coordinator, which then hands off work to other agents, and gets updates from them to report to you. 

But getting agents to genuinely share what they know—and to reach the sources they need—is where the plumbing gets trickier than the demos suggest. Kimberly says this is where careful oversight and troubleshooting are a necessary part of the process.

For instance, Kimberly was frustrated that her research agent kept handing back summaries that broke writing rules she’d drilled into her coordinator months earlier. She asked her chief of staff the status, it examined all the skills documents, and realized the rules she was communicating with that agent weren’t getting shared down the chain.

Kimberly also noticed she was never seeing articles from some of her preferred sources—something her agents did not instinctively flag for her. Again, step one was to ask her chief of staff what was going on. It told her there was a technical snag. Instead of giving up, she asked Claude if there was a different way to get what they needed, and together they came up with a workaround.

Step 5: Keep things running smoothly with a schedule

The upgrade that makes your agents feel less like another task and more like a team is putting their core work on a schedule. On Kimberly’s team, different agents have daily or weekly scheduled tasks: ​“They have their regular priorities, things they run at specific times.” 

By the time she opens her laptop, a daily briefing and potential website updates are ready for her to review.