How to keep track of several agents on one project

Run two agents on one repo and the first thing you lose is the picture. Each one keeps its work in its own window. To know where things stand you scroll back through both, and a session you closed tells you nothing. The agents are no better off: neither knows the other exists, so each plans as if it were alone.

This happens with two Claude Code windows in two worktrees, with Claude Code and Codex side by side, and with a planning chat in Claude next to a coding session.

Why a file of tasks doesn't hold the picture

A TODO.md in the repo is on a branch, so each worktree has its own copy. One agent ticking a line, or writing "doing", changes nothing in the other's copy, and nothing in the one you are reading. The file says what was planned, not who is on it now.

What holds the picture is one list outside the branches, which every agent writes to as it works, and a way to say "mine" that the list keeps.

What you see

On the board, every task shows who is on it, and done work shows whether its branch is merged. Right now has a row for each agent at work, with its branch and its task. For you has what waits on you. You don't have to watch the terminals to know which agent is doing what.

Claims

In Hamloy an agent claims a task before it starts. That is how the board knows who is on what: the claim is one session's, and it shows on the task's row the moment it is made. Any other session that reaches for that task is told who has it and since when, and given two ways on: pick another, or message the holder.

In Codex, while Claude holds the task:

• Called hamloy.claim({"key":"B-u5oh"})
  └ B-u5oh is claimed by claude since just now; pick another item (B-p2cw is free), message them with the message tool (to: B-u5oh), or pass force to take it over

The claim works across tools, because the list isn't in any of them: Claude Code, Codex, Cursor and Claude all claim on the same lists. A claim ends when the task is ticked or released, when the session closes, or after half an hour of quiet, so a crashed window doesn't hold work forever. A restarted Claude Code is told the claim is its own and takes it back.

Hearing each other

Each agent hears the others' claims, ticks and notes in the news under its next tool reply, whatever tool that is:

Hamloy news (what changed since you last looked):
- claude (session 3f9a2c71) claimed 2 items: B-u5oh "Streak counter on the home screen", B-xtiv "Push reminders at the habit's time"; leave them to them, or message them with the message tool (to: the key)

When two tasks touch the same code, the agents message each other by task key:

• Called hamloy.message({"to":"B-u5oh","text":"I'm changing HabitStore.swift for sync. Leave that file to me for the next hour?"})
  └ sent to claude (on B-u5oh); they get it under their next tool reply

With the program set up, an agent is also told when another one edits a file it edited, with the session to message.

Split the work into tasks that stand apart

A claim says who is on a task. It says less when the tasks themselves overlap. A few habits keep them apart:

• Small tasks. "Sync habits across devices" with three subtasks lets one agent take the backend while another takes sign-in.
• Say what waits on what. A task that waits on another isn't ready, so list with status: ready won't offer it to an idle agent.
• One task, one branch. Hamloy follows the work through git: a task ticked on a branch is done there, and done for everyone once merged.
• Give work to an agent. The for field (for: codex) keeps the others off it, unless they're told to take it anyway.

The same work, seen your way

The lists are one thing and how you look at them another. A view shows the same items as a list, a board or a table: who is on what, what waits on something, what was done lately. Your agents can make one when you ask.

For the whole picture, see the overview, Claims and Messages.