A new kind of job is showing up in AI, and it is not the one most people picture. It is not writing code. It is directing a small team of AI agents: deciding what each one is allowed to do, checking the work, and stepping in when something goes wrong. Companies are calling it things like agent manager or chief agent officer, and the titles are still settling.
You do not need to be an engineer to be good at this. You need judgment, a clear head and proof that you can do it safely. This guide gives you a free path to build all three: the courses, a project that proves you can run a team of agents, and a one-page scorecard that you can hand to an interviewer.
What the job actually is
An agent is an AI that does more than answer. It can take steps: search, read files, draft, check, and sometimes act. A team of agents divides a job among several of them. Somebody has to decide how that work is split, what each agent may touch, and what a human must approve first. That somebody is the agent manager.
The skills sound a lot like managing people: clear instructions, limits, review, and accountability. Which means a lot of experienced professionals are closer to this than they think.
An honest note on the numbers
You will hear that these jobs pay up to $260,000. Titles and pay vary hugely by company, and I could not trace that exact figure to a posted job. For a sense of scale, large technology companies list related AI roles with wide ranges. One posting for a principal product manager in AI sales agents listed roughly $140,000 to $275,000. Treat the headline numbers as the top of a wide range, not an average, and judge any specific role by its own listing.
Our earlier guide, The Agent Manager Playbook, covers how to translate your existing experience. This one is about building the proof.
Step 1: Take the free courses (about three hours)
Anthropic, the company that makes Claude, runs a free course library called Anthropic Academy at anthropic.skilljar.com. The courses are self-paced and free. The platform tracks your progress and completion, which is how certificates are provided, so check on each course page for the certificate.
Start with these three, in this order:
- AI Fluency: Framework & Foundations. How to work with AI effectively, efficiently, ethically and safely. This is where you learn to delegate to AI instead of just prompting it, and it is the best place to start.
- Introduction to Claude Cowork. Hands-on work with Claude on real files and projects, including managing tasks with several steps.
- Introduction to agent skills. How to build, set up and share reusable skills, the building blocks your agents will use.
If you want a gentler warm-up first, Claude 101 covers everyday use, and AI Capabilities and Limitations explains how AI works. If you run a business, there is also an AI Fluency for Small Businesses course.
Take notes as you go. You will use the words in your scorecard and your interview.
Step 2: Build the proof, a three-agent team
Courses show you understand it. A project shows you can do it. Here is a project that takes an afternoon.
Choose the tool
You can build this in Claude Code or in Codex (OpenAI's equivalent). They look and feel much alike: a chat where you describe what you want, and the tool works on it. Neither needs you to write code. If you are not ready for that, the same exercise works in Claude's regular chat, where you ask it to work through the three roles one after another. You will not get true separate agents, but you will still practice the thinking.
Choose the job
Pick a real, small job you actually need done. Good ones: researching three options for a family trip and recommending one; writing a one-page summary of a topic for your team; preparing a comparison of two vendors for your small business. Avoid anything involving money, passwords or private data for this first project.
The three agents
- The Researcher finds and organizes the facts, and records where each fact came from.
- The Drafter turns the research into a first draft.
- The Reviewer checks the draft against the research, flags anything unsupported, and lists what to fix.
I want to run a small team of three agents on one real job. Please set up and run them in this order. THE JOB: [DESCRIBE THE JOB IN TWO OR THREE SENTENCES] AGENT 1, RESEARCHER Task: find the facts I need and list the source for each one. It may: read and search public information. It may NOT: write the draft, contact anyone, or change any files outside this project. AGENT 2, DRAFTER Task: write a first draft using only what the Researcher found. It may: write and edit the draft. It may NOT: add facts that are not in the research, or send anything anywhere. AGENT 3, REVIEWER Task: check every claim in the draft against the research and flag anything unsupported, unclear or missing. It may: comment and suggest edits. It may NOT: approve the final version. Only I can. APPROVAL GATE After each agent finishes, stop and show me the result. Do not move to the next agent until I say "approved." LOG Keep a running session log: what each agent did, what it produced, what it was not allowed to do, and any time it tried to go outside its limits.
Three things that make this a real management exercise
- Hard limits. Each agent has a list of what it may touch and what it may not. Notice whether it respects them.
- An approval gate. Nothing moves forward or goes out until you say yes. This is what people mean by keeping a human in the loop.
- A rejection. Reject the first draft on purpose, and tell the Drafter why. This shows how the team responds to correction, and gives you something concrete to grade.
Step 3: Read the log and grade your team
When the job is done, ask the tool to give you the session log, and read it as a manager would. Look for: places an agent went outside its limits, facts that had no source, how the Drafter handled the rejection, and what the Reviewer caught that you would have missed.
Then fill in the scorecard below. Keep it to one page.
AGENT TEAM SCORECARD Job: [WHAT YOU ASKED THE TEAM TO DO] Tool used: [CLAUDE CODE, CODEX, ETC.] Date: [DATE] AGENT 1, RESEARCHER Did it stay within its limits? Yes / No. Notes: Were the facts sourced? Score 1 to 5: What it did best: What I would change: AGENT 2, DRAFTER Did it stay within its limits? Yes / No. Notes: Did it add anything that was not in the research? Yes / No. Notes: How did it respond to my rejection? Score 1 to 5: What it did best: What I would change: AGENT 3, REVIEWER Did it catch real problems? Yes / No. Examples: Did it miss anything I found? Yes / No. Examples: What it did best: What I would change: THE HUMAN IN THE LOOP Where did I step in, and why? What would have gone wrong without the approval gate? OVERALL Would I trust this team with a higher-stakes job? Why or why not? Three changes I would make to the limits or instructions next time:
That page is your interview artifact. It shows what a hiring manager wants to see: you set limits, you stayed in control, you measured the results and you know what you would change. Anyone can type a prompt. Very few people can show how they managed the output.
Step 4: Turn it into words an employer understands
A resume bullet you can adapt
"Designed and ran a three-agent workflow (research, draft, review) with defined permissions and a human approval gate; evaluated each agent on a written scorecard and documented improvements."
A thirty-second answer to "tell me about your experience with AI agents"
"I took Anthropic's AI Fluency and agent courses, then built a three-agent team for a real task. Each agent had hard limits on what it could touch, and nothing moved forward without my approval. I rejected the first draft on purpose to see how the team responded, then graded each agent on a scorecard. I can walk you through what I would change next time."
The vocabulary
- Agent: an AI that takes steps toward a goal, not just answers questions.
- Orchestration: deciding how several agents divide a job and in what order.
- Handoff: when one agent passes its work to the next. Handoffs are where mistakes usually creep in.
- Guardrails / permissions: the limits on what an agent may do or touch.
- Human in the loop: a person reviews and approves before anything important happens.
- Approval gate: a built-in stop where the agent must wait for a yes.
- Session log (or trace): the record of what each agent did and why.
- Evaluation (evals): measuring how well the agents did, which is what your scorecard is.
The next level
When the first project feels easy, build something that matters: an agent team that produces a weekly report for your team, triages your inbox into "needs me" and "does not," or prepares a monthly summary for a club or small business. Keep the approval gates. Add a second scorecard and compare the two. Showing improvement from one version to the next is exactly the kind of evidence that stands out.
Where this can go wrong
- Do not give agents access to money, passwords or anything you cannot undo while you are learning.
- Check the facts yourself. A Reviewer agent is a help, not a guarantee.
- The field moves fast. Tool names, menus and course titles change. Look for the closest match.
Next step
Start the first course today and block two evenings this week for the project. For more on how experience translates, read The Agent Manager Playbook, and for the thinking behind delegating to AI, see the four AI fluency skills. If you want to practice with Claude on real files, Claude Cowork: Five Things is a good place to begin.
The one line to remember
Anyone can prompt an agent. The job is proving you can run a team of them.