How to Train for the Agent Manager Role with Claude

Everyone is talking about AI agents. Very few people can actually run them well. From the outside it looks like a technical job. In practice it is a management job: deciding what to hand over, writing a clear brief, checking the work, and fixing the process when the work is wrong.

That is the skill I use every day when I build my own apps with Claude Code. Here is how to practice it in one week, with tools you can open today.

What an agent manager actually does

In February 2026, Harvard Business Review published To Thrive in the AI Era, Companies Need Agent Managers by Suraj Srinivasan and Vivienne Wei. They describe leaders who orchestrate how AI agents learn, collaborate, perform, and work safely alongside people, and they compare the role to what product managers became during the software era.

Strip away the title and the work is simple to describe. You own an outcome. Some of the steps are done by agents. You decide which steps, you write the instructions, you set the quality bar, and you are the one who signs off. The people who are good at this are rarely the most technical ones. They are the ones who already know what good work looks like in their field.

Step 1: Pick three tasks you already repeat

Start with your most repeated work, and leave your hardest work for later. A task is a good candidate when:

• It happens at least once a week.
• You can tell in under a minute whether the result is right.
• The inputs already live somewhere you can point to: a folder, a spreadsheet export, a doc.

Good first picks: a weekly status summary, sorting incoming requests, turning meeting notes into owners and deadlines, or first pass QA on a draft.

Step 2: Write each task as an agent brief

In Claude Code, a custom agent is a markdown file. Put it in .claude/agents/ inside a project to share it with that project, or in ~/.claude/agents/ to use it everywhere. The top of the file needs a name and a description. The description matters most, because it is how Claude decides when to hand work to that agent. The subagents docs list every field.

---
name: notes-to-actions
description: Turns raw meeting notes into a list of decisions, owners and deadlines. Use whenever I paste meeting notes or point to a notes file.
tools: Read, Grep, Glob
---
You turn meeting notes into an action list for the team.

Steps:
1. Read the notes in full before writing anything.
2. List decisions first, then actions.
3. For each action, give the owner and deadline only if the notes state them.

Rules:
Never invent an owner or a date. Write "unconfirmed" instead.
Keep it under 200 words.

Before you hand back, check every name and date against the notes.

You do not need to write this by hand. Ask Claude to create the file, then read it and fix anything below your standards. That reading is the management part.

No terminal? You can practice the same thing in the Claude app. Create one Project per task and put the brief in the project instructions, or package the brief as a skill.

A prompt that drafts all three briefs in one go:

I want to hand three recurring tasks to AI agents and manage them myself.
The tasks are: [task 1], [task 2], [task 3].

Before writing anything, interview me about each task, one question at a time:
what the inputs are, what a finished result looks like, what I would reject it for,
and anything the agent must never do.

Then write one agent file per task for .claude/agents/ with a short name,
a description that says clearly when to use it, the steps in order,
the output format, the rules, and a final self check.
Give me one test input per agent so I can run all three today.

Step 3: Review the work like a manager

The trap is quietly redoing the work yourself. It feels productive, and it teaches the agent nothing. Read the output and ask four questions:

1. Is it true? Check facts and numbers against the source. A claim you cannot trace is not finished.
2. Is it specific? If this output could have been written for any company or any week, the brief was too thin.
3. Is the format right? Length, structure, tone. Format problems are the easiest to fix with a rule.
4. Would I put my name on it? If not, name the one thing stopping you.

Step 4: Turn every correction into a rule

Each note you give should end up written down, so you never give it twice. If it is about one task, add it to that agent's file. If it applies to everything in the project, such as tone, formats or folders that must not be touched, add it to CLAUDE.md, which Claude Code loads at the start of every session. The memory docs explain where each file lives.

Step 5: Keep a simple log

One page, five columns: task, agent, date, what your review caught, rule added. After a week you can see quality going up and why. Anyone can say they use AI. Very few people can show how they made it reliable, and this page does exactly that.

A weekend project: a three agent research desk

Once the basics feel natural, try a small pipeline. Pick five companies you would like to work with and give Claude Code this:

Create three project subagents and run them as a pipeline.

researcher: read only. For each company, find what they sell, one recent piece
of news and one task where AI could help them. Every claim needs a source link.

writer: turns each research summary into a short, specific intro email
and a half page brief. Uses only facts from the researcher.

reviewer: checks the writer's work against the research. Flags anything
without a source and anything generic enough to send to any company.
Scores each brief from 1 to 10. Anything under 8 goes back to the writer.

Run it on: [company list]. Show me the final briefs, the scores,
and a log of what the reviewer caught. Do not send anything.

Every piece maps to real agent management. The read only researcher is a guardrail. Separate agents with separate context keep each one focused. The score threshold is a simple eval. And you reading the final log is the human in the loop.

Two guardrails before you run anything: limit each agent's tools to what it needs, and keep yourself as the approval step for anything that leaves your computer, like emails, posts or client documents.

The vocabulary that matters

• Agent: a model working toward a goal over many steps, using tools.
• Subagent: a helper started for one job, with its own context, that reports back a result.
• Orchestration: how work is split across agents, in what order, and how results come together.
• MCP: the Model Context Protocol, an open standard for connecting agents to apps and data.
• Guardrails: hard limits on tools, files and actions.
• Evals: repeatable checks that score output, so you improve on evidence.

Free material worth your time

• AI Fluency: Framework & Foundations, a free course on deciding what to delegate and how to judge the result.
• Introduction to subagents and Introduction to agent skills, both from Claude Academy.
• Building effective agents and Effective context engineering for AI agents from Anthropic's engineering blog.

Courses give you the words. The log gives you the proof. Pick your three tasks tonight, write the first brief tomorrow, and review the first output as if a new hire had handed it to you.