First fleet: a starter kit

Two AI agents in one folder: a builder does a job, a reviewer checks it. You answer their questions. This is the smallest version of how Catalyst runs its own agent fleet: one agent builds, one reviews, and the human sees only what needs a human.

The folder comes with a sample job: turn meeting-notes.txt into a one-page summary. Run it once as it is, then replace the brief and the notes with your own job.

What you need

• The Herdr app (herdr.dev, free and open source). It's a desktop app for running coding agents. They keep working when you close the window. Download it for Mac, or run the one-line installer from herdr.dev/download (macOS, Linux, Windows). It comes with hc, the command that starts and connects agents.
• Claude Code, version 2.1.224 or later. Built-in messaging between sessions needs that version. Check with claude --version.

What's in the folder

File	What it's for
BRIEF.md	The job: the goal, what done looks like, what's off limits. Write yours here.
DOORS.md	What can't be undone. The agents ask you before any of it. You decide the list.
QUESTIONS.md	Where the agents write a card when they need a decision from you.
meeting-notes.txt	The sample job's input. Replace it with your own.
prompts/builder.txt, prompts/reviewer.txt	The two opening prompts.

The agents write DONE.md (what was built) and REVIEW.md (what was checked). Those two files are your receipts.

Run it

1. Open Herdr, and open a terminal in this folder.

2. Start the two agents, one name each:

   hc run --bg --name reviewer claude -n reviewer
   hc run --bg --name builder claude -n builder

   Each name works twice: Claude Code's messaging finds the session by its -n name, and Herdr finds the terminal by its --name.

3. Open the reviewer, paste prompts/reviewer.txt, and press Enter. It waits.

4. Open the builder, paste prompts/builder.txt, and press Enter.

What to expect

• They talk by name. When the builder finishes, it messages the session named reviewer: "ready for review". The reviewer reads the brief and the work and writes REVIEW.md. Then it either tells the builder what fails or tells you nothing fails. They loop until nothing fails. Either agent can also reach the other through Herdr with hc send reviewer "...".
• Permission prompts come to you. An agent asks before it runs a command, including hc send. A message from the other agent never counts as your permission.
• You don't relay. Your job is to answer cards in QUESTIONS.md and the permission prompts.

A real run of this sample job in the Herdr app, on 8 November 2026, took about a minute. The builder messaged the reviewer through Herdr, the reviewer found nothing that fails the brief (and flagged two trivial slips that don't), and there were two permission prompts.

Make it yours

Keep the first job small: something you could check yourself in ten minutes. Write the brief so a stranger could tell when it's done. Put anything you'd regret on the doors list. When one job works, add a second.