---
title: "A lead session and a peer per project in Claude Code"
description: "One lead session keeps the ledger and checks the work while each project gets its own peer session."
url: "https://michael.zazon.us/notes/running-six-to-nine-claude-code-sessions-at-once"
published: "2026-09-22T12:00:00-04:00"
kind: "build-log"
subject: "ai-agents"
tags: ["claude"]
series: "Coding agents"
part: 3
---
I had a list of 21 things to do across a bunch of projects. Doing them one at a time would have taken so long. So I opened one Claude Code session at the root of my vault to act as the lead, while peer sessions were already running in the project repos, some of them in the middle of their own work.

Each project already has its own repo and context, so one session per project made sense, with one lead keeping the ledger and checking the work. The separation of duties is enforced at the user scoped claude.md file with instructions on how to spin up a new session in tmux if need be.

The ledger is a single vault note, each row includes the owning session and a status. By the end of that week, the status page showed nine [tmux](https://github.com/tmux/tmux) windows, one for each project. Two of them weren't Claude but [Codex](https://github.com/openai/codex) and agy, but they sat alongside the others. Underpinning it all is the machine-wide setup I described in the previous part of this series, [My Claude Code setup one chezmoi source for every machine](https://michael.zazon.us/notes/my-claude-code-setup-one-chezmoi-source-for-every-machine).

[Diagram: Lead session with ledger hands work to nine cross-checking peers, my yes in each window](https://michael.zazon.us/notes/running-six-to-nine-claude-code-sessions-at-once)

[Image: The lead session's status table in Obsidian: done this week, in flight and waiting on me for each project](https://michael.zazon.us/notes/running-six-to-nine-claude-code-sessions-at-once)

Having the specialist agent for each project be responsible for its own project or delegation to another project agent actually seems more token efficient and quality efficient because of the domain knowledge the agent has.

The sessions also cross-checked each other's work. Hours after Memento shipped its carry block code, the domains session raced two carry commands and lost the update in 60 of 60 runs. Both were writing the same temporary file, without a lock. The next Memento release had an exclusive lock, a separate temp file for each write, and a race test that failed on the old code. Would I have caught that on my own before it ate a carry block? The lead got caught making mistakes too, and caught the peers' mistakes in turn. For more on the memory tool, see [Memento agent memory I can read edit and prune](https://michael.zazon.us/notes/memento-agent-memory-i-can-read-edit-and-prune).

Claude code has agent to agent messaging now (session messaging) which has proven to be a really effective cross agent collaboration mechanism. Sessions can find each other and talk directly. This is covered with the messaging tools in the [Claude Code subagents docs](https://docs.anthropic.com/en/docs/claude-code/sub-agents).

This pattern played out in real life when I rolled out mini-splits into Home Assistant at home. I spun up three cooperating sessions (firmware/hardware, DHCP/Wi-Fi, Home Assistant adoption/cutovers) and got nine heads live in two days (see [Nine Mitsubishi mini-splits in Home Assistant for 15 per head](https://michael.zazon.us/notes/nine-mitsubishi-mini-splits-in-home-assistant-for-15-per-head) for the full build).

The rule that I approve things in each session's own window stays. But the decision rights and approval gates are a whole other post (see [Running an agent fleet like a small team](https://michael.zazon.us/notes/running-an-agent-fleet-like-a-small-team)).

Next is getting NanoClaw agents to message back and forth with host sessions. Each host session now registers its name, its tmux pane and whether it is busy, but the NanoClaw agents can't reach that registry yet. This brings my always on agent fleet into my CLI agent fleet.

_Edited on 2026-10-24 to add the mini-split rollout and the host-session registry._