One Agent, One Worktree: Isolation and Concurrency for Coding Agents

Calling parallel coding an isolation problem before a coordination problem is the right ordering. I keep separate worktrees per agent and the thing that still bites me is the merge step, because two agents can each pass their own tests and still conflict at the semantic level, like both renaming the same symbol differently. Do you gate merges on a combined test run, or only per-worktree?

I keep a tiny golden set of prompts next to every agent change. If those three break I roll back before anything else ships.

This is a fascinating shift in how we think about software development with AI. Running multiple coding agents in parallel sounds powerful, but the real challenge is not speed, it’s coordination. Giving each agent a clear workspace, responsibility, and boundaries can make collaboration much more effective. It feels similar to managing a human engineering team: the right structure often matters more than just adding more people or tools. Great insights on building a more organized future of AI-assisted development!

The isolation model here holds cleanly for anything that lives inside the repo, branch, worktree, ports, containers, db names. Where it gets harder is a task whose side effect leaves the repo entirely: it called a third-party API, posted something, fired a webhook. You can namespace a database per task, but you can't namespace a request that already landed on someone else's server. By the time a branch reaches the merge queue, verification decides whether the code merges, but for that kind of task the risky part already happened before merge, independent of what merge decides. I've ended up treating any task with a real external effect as needing its own pre-flight or dry-run gate before it's allowed to run for real at all, separate from the isolation and merge machinery, rather than trying to stretch that machinery to cover it.