Run Coding Agents in the Background
Tag a task and the factory triages, plans, builds, verifies and opens the pull request, stopping only when a human is needed. Attach to the session, answer, detach, and cap the work at three tasks at a time.
Chapters
- A factory picks up tagged tasks
- Background, not autopilot
- Attach when a human is needed
- Three tasks at a time
Transcript
A factory picks up tagged tasks
So I wanted to do like how I am doing it. so basically I tag for my software factory, factory ready, then automatically the software factory will pick my task. So I can show a little bit around my factory. So this is how the factory will pick all these tasks and it will do like triage, plan, build, verify, and it creates the PR. And in some cases, it might be like if you go and check the if I go and check the run, these are automatically shipped from the issues and these are cancelled, and this one needs human review, and I will go and update that. So I can briefly showcase what that factory look like. so these six tasks are created, and one of the issues is like this. so for that issue based on the spec, it created triage, and after the triage it created a plan, and for that plan it then executed and after execution it records some needs a human review. Then the human went like me went and type this for that, and after that it is able to pick again and it is able to pass and it is able to do that, and automatically it creates the PR.
Background, not autopilot
Now, does factory essential for for you know does creating a factory essential for you know isolating your multiple issues? No, definitely not. I created this factory in order to have you know parallel execution as well as without you know losing your mind that what is going on. Rather, it can be more executed in the background, and whenever a human is needed, you can go and do it. That way you can execute in your local machine or you can execute in your cloud environment automatically. oftentimes people misunderstood factory as well. Like when I say software factory, either the product and engineering team will go, oh no, we don't want to automate you know with the AI slop, or we don't want to automate end-to-end, how do we know the quality in the middle in a in a terminal where we can see the quality and we can stop immediately and we can go for it.
Attach when a human is needed
What I do for my factory because factory runs in an isolated environment where I don't see terminal at all, it automatically picks it calls my coding agent and it is able to plan all the five stages I shown. it goes through that. then how how does my factory works for that? I use tmux for it. like whenever a review is needed, I will go and you know check the tmux and you know attach the factory with the tmux and automatically I will I will be able to find what is needed and I will go and type it. So instead of always on, like for example, I can spin up 20 different tasks at the same time, and I can execute it in a VM environment and I can attach to the tmux with that environment, and I can go and update wherever a human is needed, human input is needed, and I can watch through the process. So, of course, it requires a lot of things to be done.
Three tasks at a time
And in my factory config, I usually say like I'm not spinning up 20 different issues. Usually I use concurrency of three, three tasks at a time, and then queue the next task once this is finished, next three and so on. so three if one is finished, then another one will be ingested. So that way it will be much more manageable for me as well in terms of context switching and so on. Still, it will be parallel rather than sequential, but without me all the time sitting and babysitting through the process.
This lesson is cut from the free Lightning session Build Your First Software Factory Execution Harness. The Live Labs build the same system with you, free, on Thursdays: see the Live Labs.