Anthropic recommends a git worktree per agent. Your runtime infra makes that a problem.
Anthropic recommends a git worktree per agent for parallel coding sessions — Boris Cherny called it the single biggest productivity unlock. The article argues the pattern is sound but your runtime infrastructure (databases, caches, shared services) turns isolated branches into a coordination nightmare. Vince and Ava land on the same insight they've been tracking since ep 770: the architecture has to match the task shape, and worktrees solve the git problem while creating a new one at the runtime layer.
Transcript
Vince Okay, so Anthropic's big recommendation for parallel agents is just… use git worktrees. One worktree per agent, each on its own branch. Boris Cherny — the guy who built Claude Code — said it's the single biggest productivity unlock from their team.
Ava Right.
Vince And I love this because it's such a boring answer. Everyone's looking for some exotic multi-agent orchestration framework, and the advice is literally learn git worktrees.
Ava It's a good pattern. The isolation is real — each agent gets its own working directory, its own branch, no file conflicts. You run five Claude sessions in parallel, each one's poking at a different part of the codebase, and they don't step on each other.
Vince Yeah, and they showed this whole thirty-PR morning thing. Like, one person orchestrating thirty-plus pull requests in a day using worktrees and MCP.
Ava Mm-hm.
Vince My week's been mostly reading stuff like this, by the way. I think I've consumed about forty articles about agent workflows and I'm starting to dream in pull requests.
Ava Oh, that's healthy.
Vince Very. So the article's argument is the worktree pattern works great for git, but your runtime infrastructure makes it a problem. That's the part I want to dig into.
Ava Yeah, and this is where it gets actually interesting. The worktree isolates files. It does not isolate your database, your cache, your message queue, any shared service. So agent one is rewriting the auth flow in its worktree, agent two is doing a schema migration in its worktree, and they're both pointed at the same Postgres instance.
Vince Right, right.
Ava The database doesn't branch with your git worktree. So your clean isolation at the file level completely falls apart at the runtime level, and now you have two agents corrupting each other's assumptions about what the schema looks like.
Vince This is the ep seven-seventy thing all over again. We said decomposition isn't universally safe — if the subtasks are separable, parallel works. If they share state, each handoff adds drift.
Ava Exactly. And worktrees are the git layer's answer to decomposition. They're betting the tasks are separable. But the runtime layer is where that bet gets tested, and most cloud-native apps have a fat shared-state layer that doesn't care about your branch structure.
Vince So who's this actually for? Like, if I'm running agents on a monorepo where each agent touches a different package and there's no shared database in the loop…
Ava Then you're fine. Genuinely. The pattern works. Spin up your five worktrees, let them rip.
Vince Okay.
Ava The problem is when your definition of done includes running the thing. If the agent needs to start the server and hit an endpoint to verify its changes, and that server depends on a shared database, you need a database per worktree. And now you're talking about provisioning five Postgres instances per coding session, and nobody's built that infrastructure.
Vince That's the product gap, right? The git tooling is solved. Worktrees have been in git forever. But per-agent runtime isolation — ephemeral databases, disposable environments that branch with your code — that's where the infrastructure hasn't caught up to the workflow.
Ava Agreed. And the article's right to call this out. The worktree advice is correct but incomplete. It solves the easy layer and punts the hard one.
Vince Oh, come on.
Ava What?
Vince Nothing, that's just such an Exploring Next sentence. The easy layer's solved, the hard layer's unsolved, and we're all pretending the advice is complete.
Ava Well, it's true.
Vince No, it is. I'm not even arguing. It's just funny how every one of these articles ends in the exact same place. The interesting work is always one layer below where the recommendation stops.
Ava That's because the easy layer is where you can write a tip on X. The hard layer is where you need provisioning and lifecycle and cleanup, and that doesn't fit in a tweet.
Vince So if you're actually doing this today — worktrees plus parallel agents — what do you do about the runtime problem?
Ava You either keep your agents on truly independent code paths with no shared state, or you build the ephemeral environment yourself. Testcontainers, per-worktree database snapshots, something that gives you isolation at the runtime layer. The article doesn't prescribe a solution, which is honest, because there isn't a clean one yet.
Vince Ask us in six months, basically.
Ava Yeah. Ask us in six months.