How to manage multiple coding agents
Running three or four AI coding agents is easy. Knowing which one needs you is the hard part. A working method, and where the attention goes.
By Piotr Jura, creator of AgentNotch. Tested on macOS with AgentNotch v1.3.1.
Starting three agents is easy. You give each one a task, and for about four minutes it feels like the best day you have had at work. Then the first one finishes, and nothing tells you.
That is the whole problem, and it is not a tooling problem at the start — it is an attention one. A finished agent is silent. A stalled agent is silent. An agent halfway through a long tool call is silent. From outside, the three states are one state, so you go looking, and each check costs a window switch and a re-read of where you were. Do it often enough and the parallelism you bought is spent on supervision.
A method that holds up
One task per agent, one worktree per task. Give each agent its own checkout so two of them can't rewrite the same file. Worktrees are cheap and they end an entire category of problem.
Write the task down before you start it. Agents drift when the brief is thin, and a drifted agent is the most expensive kind: it works confidently for ten minutes in the wrong direction.
Decide who you check first — before you need to. Without an order you check the agent you last touched, which is the one least likely to need you.
Watch the limits, not just the work. Four agents spend a shared plan several times faster than one. Running out mid-task is a self-inflicted interruption.
Where the attention actually goes
The method above removes the collisions and the drift. What it can't remove is the checking. That is the part worth automating: something that watches the sessions themselves and tells you which one stopped, which one is asking a question, and which one is waiting for approval on a command it doesn't want to run alone.
Where AgentNotch fits
AgentNotch is that layer, in the macOS menu bar. Every agent you run — Claude Code, Cursor, Codex, Kimi, and Cowork sessions from the Claude app — appears with its status, branch, model and context, and the one that needs you is flagged. Approvals and questions can be answered from the notch. It reads the agents themselves, so it works the same in Ghostty, iTerm2, Warp or Conductor's worktrees, and everything except plan usage stays on your Mac.
Side by side
| The problem | What actually fixes it | |
|---|---|---|
| You forget an agent is waiting | A finished agent is silent, and silence looks like work. | One place that says which one stopped. |
| You check the wrong window first | Round-robin scanning costs more attention than the work. | Status you read without switching. |
| Two agents edit the same files | One task per agent, one worktree per task. | A branch shown next to every session. |
| You run out of plan mid-task | Parallel work spends a shared limit several times faster. | Plan limits visible while you work. |
Questions
- How many coding agents can you realistically run at once?
- Most people top out around three or four, and the limit is rarely the machine — it is how many things you can hold in your head. The cost of a fourth agent is not CPU, it is the attention you spend checking whether any of the other three stopped.
- What is the hardest part of managing multiple AI coding agents?
- Knowing which one needs you. A finished agent, a stalled agent and an agent thinking hard all look the same from outside — a quiet window. Everything else about running several agents is easy by comparison.
- How do you stop agents from colliding on the same files?
- One task per agent, and a git worktree per task, so each agent has its own checkout. It costs a minute to set up and removes the entire class of problems where two agents rewrite each other's work.
- Do I need a dashboard to run several agents?
- You need somewhere that answers 'who needs me?' without a window switch. That can be a menu-bar app like AgentNotch or a discipline of checking in a fixed order — the failure mode is having neither and discovering an idle agent twenty minutes later.
Keep reading
See which agent needs you.
Try AgentNotchfree for 14 days, no card. Watch Claude Code, Cursor, and Codex from your Mac's menu bar.
Updated 2026-07-26