You can now orchestrate subagents running in Claude Code, Codex, or the Warp Agent. Let the agent create a delegation plan, and watch subagents coordinate with message passing. Each agent gets its own worktree for isolation. Here's how it works with eng Matthew Albright:

Claude Code and Codex subagents are now in beta. Try it out in the latest version of Warp by running /orchestrate with your next Warp Agent prompt!

Delegation plan + isolated worktrees is the right combo. Most multi-agent setups fail because agents share state and step on each other. Giving each one its own environment turns it from a coordination nightmare into real parallelism.

per-agent worktrees is the kind of boring correct call that everyone copies in 6 months after their shared-state setup melts down

That's amazing thank you

The worktree isolation per subagent is the key design decision here. Most multi-agent setups break down because of shared mutable state — agents stepping on each other's changes. Giving each subagent its own isolated context is what makes the delegation model actually work at scale.

@zachlloydtweets How does orchestration approach prompt caching? Does the main agent build the initial context like planning mode, and then fork its context to create the delegates?

I'd definitely switch to Warp if it were as competitive in tokens/price as Claude Code Max 5x

subagents with isolated worktrees solve the parallel coordination problem. the leftover one is session loss, every parent and child still dies on a Mac restart and the message-passing context evaporates.

Feels like agent orchestration is becoming its own engineering layer now. The hard part is less about generating code and more about task decomposition, coordination, and preventing context drift between agents.

When two isolated worktrees return conflicting patches, who owns the conflict at merge time?