Train for the agent manager era: free courses, one weekend build, the vocabulary

The job, with receipts 🎓

• 📰 Harvard Business Review, February 2026: To Thrive in the AI Era, Companies Need Agent Managers, co-written by a Harvard Business School professor. The role: people "responsible for orchestrating how AI agents learn, collaborate, perform, and work safely alongside humans." Their comparison: agent managers are to the AI era what product managers were to the software era
• 🏢 It already exists. The same article profiles real Salesforce agent managers running fleets of AI support agents ("I start and end my day in dashboards, scorecards, and agent observability")
• 📈 Demand is measurable. Microsoft's Work Trend Index: 28% of managers are considering hiring AI workforce managers and 32% plan to hire AI agent specialists within 12 to 18 months. McKinsey: demand for AI fluency grew sevenfold in two years, faster than any other skill in US job postings
• 🤫 And the gap that makes it valuable: an HBR survey found only 6% of companies fully trust AI agents to run core processes on their own. Someone has to be the trust layer. That's the job

Your first week as an agent manager 🗓

Courses and certificates are useful, but nobody hires an agent manager off a certificate. They hire the person who is already doing it. HBR's own framing is that domain expertise matters more here than being technical, which means the qualification is the thing you already have: you know what good work looks like in your job. These five moves turn that into evidence, using one week and tools you can open today.

1. 🗂 Pick the three tasks you already repeat. Not your hardest work, your most repeated work. A task qualifies if it happens at least weekly, has an output you can recognise as right or wrong in under a minute, and the inputs live somewhere you can hand over (a folder, a doc, an export). Good starters: the weekly performance report, inbox and request triage, turning meeting notes into owners and deadlines, repurposing one long piece into five short ones, first-pass QA on someone else's draft.
2. 📝 Give each task its own instructions file. This is the actual craft of the job. In Claude Code, an agent is a markdown file: put it in .claude/agents/ to share it with a project, or ~/.claude/agents/ to have it everywhere. It needs a name and a description at the top (the description is what makes Claude reach for it on its own), then your instructions underneath. You do not have to write the file by hand: ask Claude to write it and it creates the file for you. Note that /agents no longer opens a wizard in current versions, it points you at the files instead. Full field list in the subagents docs.
3. ▶️ Run all three, then review like a manager. The trap here is quietly redoing the work yourself. Don't. Read the output, score it against the rubric below, and write your notes as feedback to the agent instead of edits to the document. Managing means the fix goes into the process, not into this one deliverable.
4. 🔁 Turn every fix into a rule. Each correction gets written back into an instructions file so you never give the same note twice. If the note is about how that one task should be done, it belongs in that agent's own file. If it applies to everything in the project (tone, formats, what never to touch), it belongs in a CLAUDE.md, which loads at the start of every session. The memory docs cover where each file lives and which one wins when they disagree.
5. 📓 Keep the log, because the log is the proof. One page, five columns: task, agent, date run, what your review caught, the rule you added. After a week you have a real orchestration log showing quality going up over time. That page is your LinkedIn post, your interview artifact, and the thing you take to your manager when you ask to own this properly.

To get all three agents drafted in one go, paste this and answer its questions honestly. Vague answers here produce vague agents.

I'm training myself for the agent manager role, so I want to stop doing three tasks by hand and start managing them instead. The three tasks are: [task 1], [task 2], [task 3].

For each one, interview me first: ask what the inputs are, what the finished output looks like, what "good" means to me specifically, and the three mistakes I would reject it for. Ask all of that before writing anything.

Then write me one agent instructions file per task. Each file needs a short lowercase-hyphenated name, a description line that says plainly when this agent should be used, and instructions written the way I would brief a new hire: the steps in order, the format of the output, the quality bar, and an explicit list of what it must never do. End every agent's instructions with a self-check the agent runs before handing work back to me.

Finally, tell me exactly where to save each file so it loads automatically, and give me one test input per agent so I can run all three today.

How a manager reviews, in four checks 🧠

The free course path (one afternoon) 📚

Anthropic's course library is called Anthropic Academy. Every course is free, needs only an email sign-up, and gives you a shareable completion certificate. The path, in order:

1. 🧠 AI Fluency: Framework & Foundations (about 1 hour). The 4D framework: Delegation, Description, Discernment, Diligence. This is literally the mental model for directing AI instead of just prompting it, and it has an assessment-backed certificate
2. 💼 Introduction to Claude Cowork (short). Hands-on with the agentic side of Claude where workplace orchestration actually happens: task loops, skills, plugins, integrations
3. ⚡ Introduction to agent skills (about 30 minutes), then chase it with Introduction to subagents (about 20 minutes) for true multi-agent delegation

The whole path is roughly 3 hours of video. One focused afternoon, not a degree.

The weekend build: run a 3-agent research desk 🖥

Courses teach vocabulary. This build proves you can do the job. It works in Claude Code, or in Claude Cowork if you prefer no terminal (both need a paid Claude plan, Pro is enough). Pick 5 companies or creators you'd want as clients or employers, then paste:

I want you to work like a team I manage. Create three subagents:

1. researcher: given a company name, find what they sell, one recent piece of news, and one specific marketing task AI could speed up. Read-only research, return a tight summary with a source link for every claim.

2. drafter: take the researcher's summary and write a 120-word personalized outreach email plus a half-page account brief in my voice.

3. qa-reviewer: check the drafter's work against the researcher's facts. Flag anything unsupported by a source, flag anything generic enough to send to any company, and score it 1-10. Nothing reaches me below an 8.

Then run the full pipeline on these 5 companies: [YOUR LIST]. Show me a final report with all 5 briefs, each QA score, and a log of what each agent did and what QA caught.

Why this build is the proof: you're not prompting, you're delegating, setting a quality gate, and reviewing an orchestration log. Screenshot the report showing what your QA agent caught before it reached you. That screenshot is your LinkedIn post and your interview artifact.

And every design choice maps to a term you just learned: the researcher is read-only (guardrails), each agent works in its own context window (context management), the "nothing below an 8" rule is an eval, and you reviewing at the end is human-in-the-loop.

The 10 words that get you into the room 🗣

• 🤖 Agent: AI that works toward a goal over many steps using tools, not just answering one prompt (Anthropic's definition)
• 🧩 Subagent: a worker agent spawned for one job, with its own context, reporting only the result back (docs)
• 🎼 Orchestration: deciding how work splits across agents, in what order, and how results merge (Anthropic's own research system uses an orchestrator-worker pattern)
• 🤝 Handoff: passing a task plus the context the next agent or human needs. Most multi-agent failures are bad handoffs
• 🔧 Tool use: an agent taking real actions (searching, writing files, calling apps) instead of generating text
• 🔌 MCP: the open standard that plugs agents into your apps and data, like USB-C for AI (modelcontextprotocol.io)
• 🚧 Guardrails: hard limits on what an agent may do: which tools, which files, what it must never touch
• 🙋 Human-in-the-loop: checkpoints where a person approves before the agent proceeds
• 📊 Evals: repeatable tests that score agent output, so you improve on data instead of vibes
• 🧠 Context management: controlling what each agent knows; separate context windows keep agents focused (Anthropic on context engineering)

Where the actual jobs are 💼

The title on the posting is rarely "agent manager". Search these instead: AI operations manager, AI operations lead, marketing AI operations, AI enablement lead, agent operations, AI orchestration specialist. Live examples (postings expire, so treat these as proof of the market):

• 💰 Airtable, Senior Manager, Marketing AI Operations: $184,000 to $260,600 (posting)
• 💰 Arcadia, AI Operations Lead, remote: $127,500 to $228,400 (posting)
• 📰 Forbes rounded up 20 agentic AI job titles Box, McKinsey and LinkedIn see coming, including McKinsey's "agent supervisor"

What to say to your boss (or in the interview) 💬

• "Microsoft found 46% of leaders already use agents to fully automate workstreams, and 82% expect to use digital labor within 18 months. I'd rather be the person who manages the agents than the person who waits."
• "McKinsey puts $2.9 trillion a year of US value on the table by 2030, but only if companies redesign workflows around people plus agents. Someone has to do that redesign."
• "88% of companies use AI, but only 23% are scaling agentic systems. The bottleneck isn't the tech, it's people who can operate it."

If you want to go deeper (still free) 🎁

The links 🔗

🎓 Anthropic Academy: anthropic.com/learn
📰 The HBR article: To Thrive in the AI Era, Companies Need Agent Managers
🤖 Claude: claude.com (the weekend build runs in Claude Code or Cowork, from the Pro plan)