What a Managed Agent is
Full Lesson Reference
A Managed Agent is Claude Code running without you at the keyboard. Same Claude, same skills, same MCPs, same CLAUDE.md - except nobody's watching. It fir es on a schedule, does the work, logs the result.
This module is about a different shape of Claude. Modules 1 through 9 taught you how to work with Claude in a session. This module teaches you how to make Claude work for you when you're not there.
The three shapes of Claude
- A session
You open the terminal, type claude, and work together in real time. You prompt, Claude responds, you review. This is what Modules 1 through 9 covered.
- A skill
A reusable workflow packaged into one command. You still trigger it in a session, but the steps are pre-written. Module 9.
- A Managed Agent (also called a Routine or a Trigger)
Claude running on Anthropic's infrastructure, on a schedule you set, without a local session. It has its own context, loads its own skills and MCPs, does the work, writes the output somewhere you can read it later.
What runs where
The mental shift most people miss: a Managed Agent runs remotely. Not on your laptop. Not in your terminal. On Anthropic's servers, in a fresh context that doesn't see your local files unless you give them to it.
That changes what you need to set up
- Instructions - what the agent should do, in the prompt you attach to the Routine
- Context - any files, data, or references it needs, bundled into the Routine
- Credentials - API keys and MCP connections it'll need, stored as secrets
- Output - where it should save results (GitHub, Supabase, Slack, email, etc.)
A session borrows from your local setup. A Managed Agent has to carry everything with it.
Why managed agents matter
Sessions are fine for work you can sit through. Managed Agents handle the work that shouldn't need you at all:
- Nightly data pulls from ad platforms into Supabase
- Weekly client reports emailed out every Monday 7am
- Daily health checks that post a summary to Slack
- Monthly audits that drop an HTML dashboard into GitHub
- Long-running backfills that would hog your terminal for hours
The pattern: something repeatable, on a clock, where the output is more valuable than your attention while it runs.
When a Managed Agent is the wrong choice
- One-off tasks - the setup cost isn't worth it
- Tasks that need your judgement mid-run - agents can't pause and ask you
- Anything that writes to production without review - never let an agent push ads live, send client emails, or mutate shared systems unattended unless you've tested it to death
- Tasks that change every time - if every run needs different instructions, stay in a session
The stack underneath
A Managed Agent uses the same building blocks as a session
- CLAUDE.md still loads (you bundle it as context)
- Skills still work (you install them into the Routine)
- MCPs still connect (you register credentials as secrets)
- Memory layer (Supabase) still reads and writes if you wire it up
Everything you've learned carries over. The only new thing is the runner - the mechanism that fir es Claude without you pressing enter.
Action items
☐ Understand: a Managed Agent = remote, scheduled Claude with its own context
☐ Make a list of 3 repeatable tasks in your week that shouldn't need you watching
☐ For each, note: what does it produce, where does the output go, who reads it
Next lesson: When to use a Routine.
Exercises
- Review the concepts covered in this lesson: What a Managed Agent is.
- Write down your key takeaway from this lesson.
- Practice running any commands or prompts mentioned above inside your terminal.