/batch — 5 to 30 Parallel Agents, Each Opening a PR
/batch takes a single instruction and fans it out into a swarm of independent parallel agents — each in its own worktree, each opening its own PR. It is purpose-built for large-scale parallelizable changes that do not require inter-agent coordination.
Quick Reference
- →/batch spawns 5–30 agents simultaneously, each in an isolated git worktree
- →Each agent opens its own PR — you review and merge individually
- →Shipped February 27-28, 2026 — available in all current Claude Code versions
- →Best for: bulk migrations, applying a pattern across N files, parallel feature scaffolding
- →Each agent gets branch: claude/batch-{id}
- →/batch uses worktree isolation (not agent teams) — no inter-agent communication
- →Calculate cost before running large batches: N agents × per-task cost
- →When /batch beats agent teams: work is truly independent, file-level changes, individual PR review needed
What /batch Does
/batch has three phases: (1) Research & Plan — Claude explores what needs to change and decomposes work into independent units; (2) Parallel Execution — one background agent per unit, all launched simultaneously; (3) Progress Tracking — a status table with PR URLs from each agent.
/batch is not an agent team. It does not use the Shared Task List or peer communication. Each agent is fully independent — it receives a task, works on it in its own worktree, opens a PR, and exits. The agents never talk to each other.
| /batch | Agent Teams | |
|---|---|---|
| Inter-agent communication | None | Shared Task List + direct messaging |
| Work type | Independent units | Sequential or collaborative |
| Output | N separate PRs | One synthesized result |
| Coordination overhead | Minimal | Significant |
| Best for | Bulk changes, migrations, parallel scaffolding | Sequential handoffs, specialization, peer review loops |