how many parallel agent sessions can you actually run before losing track of what is happening?

by•

​tried running three different agents on separate branches today one building the UI, one setting up the database schema, and another writing basic tests. instead of moving three times faster, my brain completely fried trying to keep track of which terminal window needed approval next.

​do you stick to one focused agent at a time or have you figured out a clean way to juggle multiple sessions?

9 views

Add a comment

Replies

Best

The approval juggling is the part I'd attack first. Our project's agent rules list what an agent does without asking: building, running, tests, generators, routine deploys. Only two things need a word from me: destroying data, and anything that reaches real users. Fewer things to approve means fewer windows to watch.

The other cost of parallel sessions showed up in git, not in my head. Several sessions share one working tree, so each one stages only the files it changed and commits by path. A bare commit once took another session's staged rename and deletion along with it and left main unbuildable until the next commit. The rule is now in the project file every session reads.

I run more than three, but not by watching each terminal. I built my own terminal for it, with up to eight panes on one screen. I don't approve every step. Every session works from the same global rules and keeps its own memory and worklog, and it only stops for me on things like a deploy, an email to a real customer, or anything that can't be undone. Like running a company, I don't need to know everything they do, just the key points.