← ClaudeAtlas

dispatching-parallel-agentslisted

Use when 3+ tasks are independent — no shared files, no order dependency — and can run concurrently. Fan-out only; per-task review happens after all return.
felipemelendez/llm-orchestrator · ★ 3 · AI & Automation · score 69
Install: claude install-skill felipemelendez/llm-orchestrator
# Dispatching parallel agents (fan-out) Send N agents in one batch. Collect N returns. Then review. ## The isolation invariant — read this first **No two agents ever write the same file at the same time, and no undeclared writers ever share a tree.** Concurrent writes to one checkout race and clobber each other (a parallel agent running `git stash`/`reset`/`add` on the shared tree once silently destroyed in-flight work). So parallelism is allowed in exactly three safe shapes: - **Read-only fan-out** — reviewers, explorers, researchers. They only read, run tests read-only, and report; they **never** edit files or mutate git. Safe to run together on the shared checkout. - **Isolated writers** — each writing agent works in its **own git worktree** (a separate directory on its own branch), so its edits and commits cannot touch the shared tree or another agent's. You merge the branches back **sequentially** afterward. This is the default for parallel writers. - **Declared shared-checkout writers** — only when the project rules out worktrees (an owner wants one visible tree, or waves stack on uncommitted sibling diffs a private worktree cannot see). Concurrent writers share the checkout under an explicit, controller-partitioned file-ownership discipline — see the shared-checkout steps below. Never run two writers concurrently on one checkout *undeclared*. If you can't isolate the writers and won't declare a shared-checkout partition, run them sequentially (`dispatching-subagen