One Manager, Many Agents

I run several Claude Code sessions at once: games, the libraries they share, a dashboard. For a while the hardest part of my week was not the code they wrote. It was me, carrying decisions, questions and approvals between their windows.

This guide is the method I use instead. One more session, the manager, keeps the records in plain files, writes the specifications the other sessions implement, checks what they report, and puts every open question to me in one place, with the options and a recommendation. It never decides for me, and it never passes on an approval. The guide explains the method part by part, then shows a full week of it at an invented studio, recorded from real Claude Code sessions.

Who it is for

You already use Claude Code every day, you have more than one session going, and you are comfortable with Git, a terminal, plain-text files and a little Python.

It is not a first-steps book, not a hosted platform, and not a way to let agents decide unsupervised.

What it gives you

• One place for everything that needs you. Questions, choices and approvals from every session reach you in the manager's window, each with its options and a recommendation, and on one local dashboard page. You stop going from window to window to find out what is waiting.
• Decisions that travel, approvals that stay put. A choice you make once reaches the session that needs it, in your own words. Anything that publishes, deploys or costs money is still approved by you, in the session that does the work. A message from the manager is never your yes.
• Reports that are checked before they are believed. The manager checks a session's report against the repository before it records it. In the recorded week it found a report claiming a lockfile that did not exist, and its release audit, run against the build rather than the records, found gaps the records did not show.
• Records you can read without an agent. Status, dated decisions with your own words, and one file per project with an ID for every open item, all in plain Markdown under Git. A fresh session starts where the last one stopped.
• Spending proposed before it happens. A paid API call is put to you first: which model, how many calls and why, the estimated cost and what it buys. Nothing runs until you say yes in that session.
• Client work kept apart. A fence between groups of sessions blocks messages between your own products and a client's work, and each group's sessions get their own rules at the start of every session.

What is inside

• The guide (PDF): the method part by part, a full week at an invented studio recorded from real sessions, the rules learned the hard way, and the templates to build it all by hand.
• The guide with the kit (PDF and zip): the same guide, plus the working parts, tested: the message fence, the session rules, the manager's workspace, the launcher and the dashboard, with a demo studio to try them on.
• What you need: Claude Code with named sessions and messages between sessions; Python 3.9 or later; macOS and iTerm2, or tmux; Codex CLI only for image runs.

Honest limits

Parts of the method lean on Claude Code behaviour that is not documented, so every claim about a tool in the guide carries the version and the date it was checked. Several sessions cost more than one. The tools are tested on macOS. The guide has a whole chapter on its weak spots.

Coming soon on Gumroad

By Michael Shvets.

The story behind it

Not affiliated with, or endorsed by, Anthropic or OpenAI. Claude Code and Codex are named to describe the tools the method uses.