Agent Workflows — Many minds, one task

Many minds,
one task.

Split the work. Run it all at once. Ship a stack.

Agent workflows take a large job — a design doc, a feature plan, a research question — and fan it out to specialized subagents running in parallel. Each worker gets an isolated git worktree, a scoped assignment, and a structured result to hand back to the orchestrator. Built for Claude Code agent teams.

01 · Fan-out

DAG from a plan

A design document becomes a directed acyclic graph of PR-sized chunks. The orchestrator topologically sorts dependencies before anything runs — independent nodes launch together.

02 · Isolation

Worktree per agent

Each implementer works in its own git worktree — no stomping on shared files. Commits land as discrete SHAs the orchestrator fetches and assembles into a branch stack.

03 · Assembly

Stack + review

Branches stack in dependency order. Mandatory reviewer subagents run before merge-ready. Output is a Graphite PR chain, not a pile of loose diffs.

How a job moves through it

Parse the planBreak a design doc into nodes with explicit depends-on edges and persona assignments — implementer, reviewer, verifier. Each node gets a scope, acceptance criteria, and a target worktree name.

Launch in parallelIndependent nodes spawn simultaneously as Claude Code agent teams. Blocked nodes wait on parent commits before their worktree is seeded — no race on shared state.

Review-fix loopEach node gets a reviewer subagent. Failures return to the implementer until the orchestrator marks the branch merge-ready. SHAs accumulate in the worktree lane.

Assemble the stackWorktrees tear down, branches fast-forward in dependency order, and PRs open — one coordinated Graphite stack instead of serial handoffs.

Coordination primitives

Three ways Claude Code composes agents — each maps to a different scope and lifecycle in the orchestrator.