← sources

I let 11 Claude Code sessions run jobs on my other machines and message each other. Here's what they did over the last 4 days.

source: reddit (reddit:ClaudeCode)published 2026-11-13 17:46ingested 2026-11-14 02:41

retained text

I run one or more Claude Code sessions per repository, and for months they kept failing in the same places. A session could not reliably run anything on another machine, and it could not talk to the session in the next repository. I was the SSH babysitter and the copy-paste relay. So I built a tool for both, called ax, with Claude Code (and some Codex) doing the building. This is what four days of real use looked like.

# Running jobs on other machines

My builds and tests run on machines that are not my laptop: a few servers, a Windows box on my home LAN and a Mac that is often on a 4G hotspot. Over SSH, Claude Code would start a build, hit a tool timeout, and then write a wait loop to find out what happened.

With ax, the session submits the command as a job and gets a job ID back. The remote machine supervises the process and keeps its status and output. When the session times out, or I close it and open a new one the next morning, the agent asks for the same job by ID and reads the output from where it stopped. It cancels by ID too, so it stops the job it meant to stop.

What did I do with this? For my remote desktop product I need to debug issues when network conditions are bad or lossy; traffic shaping can simulate it but it's not the same as running it through a real 4G mobile link. With ax, I put the test machine on a real 4G hotspot and my agents still ran their tests on it and on the other machines, without needing to add a VPN.

The machines connect to each other over a P2P connection (Iroh), so this needs no VPN or Tailscale and it works when the laptop and the build machine are on different networks.

# Sessions messaging each other

Each session registers a mailbox. Then I tell one agent "ask the reviewer to check this", and it sends a signed message. The other session is notified, fetches it and replies. Over four days, 11 sessions in 8 repositories exchanged 300 messages, about 98,700 words that I did not copy between windows. Half were picked up within 4 seconds, and half of the replies were sent within 77 seconds of that.

What they used it for:

* An implementer and a reviewer in the same repository exchanged 131 messages. The reviewer rejected or held about a third of the steps, including one run reported as complete that the logs showed was void.
* An application's agent sent change requests to the agent that maintains its component library. Seven library releases went out in 31 hours, and two requests were answered with "*the library already does this, here is how*".
* A library's agent announced each release to the projects that depend on it. One of them reported a defect back, and the fix was confirmed 18 minutes later.
* The two halves met as well. One project's agent was running its test gates as ax jobs on the Windows machine, hit failures, and messaged the ax repository's agent with the failing command and its evidence. They worked out which part was ax's fault in ten messages.

# What I learned

* A second session that only reviews is worth far more than I expected. It reads the logs and the code, and it does not accept "done" without evidence. I can watch them side-by-side if I’ve the time and it is such a relief not to be in the middle!
* Agents write better reports for each other than I ever relayed. 132 of the 300 messages point at a file the other agent can open, and 61 cite a commit.
* I still make the decisions. Releases waited for my approval, and more than a third of the messages refer to something I asked for or decided. However, I have more time now and can choose to use it personally or to get more done. **This has been the biggest single benefit for me**.
* Giving a remote command an ID changes how an agent cleans up. It checks, waits for or cancels that one job, and the process-name searches disappear.
* Agents have their own habits. They introduce themselves at the top of every message although the sender is already authenticated, and one apologised in a follow-up for a message it had sent by mistake.

ax is free and will remain free. I have some ideas for monetisation. Source is currently not available. It works with Claude Code, Codex and opencode. You install it on macOS or Linux (or inside WSL), and the machines it runs jobs on can be Linux, macOS or Windows (both native and WSL).

* Landing page: [https://ax.getaster.now/](https://ax.getaster.now/)
* Docs: [https://ax.getaster.now/docs/](https://ax.getaster.now/docs/) and full write-up [https://ax.getaster.now/docs/guides/messaging-in-practice.html](https://ax.getaster.now/docs/guides/messaging-in-practice.html)
* Problems and issues: [https://github.com/GetAster/ax](https://github.com/GetAster/ax)

Happy to answer questions!

document fetched 2026-11-14 00:44 · canonical https://reddit.com/r/ClaudeCode/comments/1xveidw/i_let_11_claude_code_sessions_run_jobs_on_my