Agent Orchestration — Parallel Builds for Claude Code Sub-Agents
Agent Orchestration — Parallel Builds for Claude Code Sub-Agents — from InitRepo, the AI project planning document generator.
Parallelization waves
We read your roadmap, user stories, and architecture, infer the dependency graph, and group the work into waves — sets of tasks with no unmet dependencies that can be built at the same time.
Ready-to-run agents
Each track becomes a real .claude/agents/*.md sub-agent file, scoped to its coordinates and files. Unzip into your repo and Claude Code picks them up immediately.
Loop recipes
A library of agentic loops — implement→verify, TDD, typecheck/lint, build-fix, spec-adherence, wave-integration — parameterized with your project's actual build and test commands.
From a plan one agent reads to a plan a fleet of agents builds
InitRepo already turns your idea into a complete planning suite — business analysis, PRD, UX/UI spec, technical architecture, user stories, and a build roadmap, all tied together by a coordinate system an AI agent can navigate. Agent Orchestration takes the next step: instead of one agent working through the roadmap serially, it produces an executable plan for several Claude Code sub-agents working in parallel.
How it works
- 1. Analyze. We feed your completed docs to the model, which extracts the concrete build tasks, their dependencies, and the files each one is likely to touch — bounded by your architecture's natural component boundaries.
- 2. Layer into waves. Code — not the model — topologically orders those tasks into waves and flags file-ownership conflicts (two tasks editing the same file in one wave), the number-one cause of parallel-agent merge conflicts.
- 3. Generate the artifacts. You get an
ORCHESTRATION.md— dependency graph, waves, copy-paste kickoff prompts, integration guidance, and the loop library — plus the.claude/agents/*.mdfiles, all bundled into your project ZIP.
What lands in your ZIP
ORCHESTRATION.md with a Mermaid dependency graph, wave-by-wave kickoff prompts grounded in your project's coordinates, a sub-agent roster, and merge guidance — alongside one ready-to-run sub-agent file per track. Each agent is told exactly which coordinates and files it owns, which files to leave alone, and to run your typecheck and tests before it finishes.
Who it's for
Builders who run Claude Code sub-agents and want to parallelize a real build without the agents stepping on each other. Agent Orchestration is a Max-plan feature — generate it on demand from any completed project, and regenerate any time the plan changes.