What is a git worktree and why is it important for orchestration?

A worktree is a working copy of the repository in a separate directory, linked to a different branch. It allows multiple Claude instances to work on the same project without interfering with each other — each in its own space, on its own branch. At the end, you merge the work. Without worktrees, parallel agents would overwrite each other's files.

How many parallel instances are optimal?

Diminishing returns generally kick in at 2–3 parallel instances. More than 3 simultaneous instances increases coordination cost (yours and the tokens) without proportional productivity gain. The exception is projects with genuinely independent workstreams — frontend, backend, tests — where 3–4 agents do produce clear gains.

How do I divide tasks between agents effectively?

The basic rule: each agent needs a clear objective, enough context to work autonomously, and access only to the files it needs. The best divisions are by layer (frontend/backend), by feature (auth/payments/notifications), or by phase (implement/test/document). Avoid divisions where agents need to edit the same files.

Is Agent Teams ready for production?

Agent Teams is experimental at the time of writing. It works well for research and development use cases where coordination overhead is acceptable. For production projects with deadlines, the recommendation is sub-level 2 (manually managed worktrees) until the feature matures.

Does Level 6 require advanced git knowledge?

It requires comfort with basic git concepts (branches, merge, conflicts) and the ability to resolve merge conflicts. It doesn't require being a git expert. The --worktree command is straightforward; the most complex part is usually managing the final merge when multiple agents have worked on adjacent areas.

Level 6 — The Orchestrator

13 March 2026

Mindset: «You're not a coder — you're a manager»

Level 6 is the top of the programme. The mindset shift is radical: you stop executing and start managing. You're not the person doing the work — you're the person who decides what work gets done, who does it, and when it gets integrated.

You manage sub-agents the way you'd manage a real team. You assign them specific tasks, give them enough context to work autonomously, and oversee the integration of the result. AI is not your tool — it's your team.

You're here when...

You run multiple Claude sessions in parallel on the same project
You use worktrees for isolated, conflict-free workspaces
You decide WHAT gets done — not HOW it gets done
You manage sub-agents and agent teams like a real team lead
You merge parallel work at the end of each cycle

Skills to master

Task division + planning parallel workstreams (auth / frontend / payments simultaneously)
Git worktrees (--worktree / -w flag) — each agent gets its own «desk», merge at the end
Multi-terminal session management and context switching
Agent Teams (experimental) — sub-agents that talk to each other + a supervisor agent that coordinates and reports to you

The 4 sub-levels of orchestration

Sub-level 1 — Multiple terminals, same branch: Several Claude Code instances on the same project. Simple, but risks conflicts if both edit the same files simultaneously.
Sub-level 2 — Git worktrees: Each instance works in its own branch/worktree. No conflicts. Merge at the end. More technical, cleaner. Start with: claude --worktree frontend, claude --worktree backend.
Sub-level 3 — Sub-agents via worktrees: Ask Claude to spawn sub-agents in separate worktrees automatically. Less visible, more abstracted — Claude manages the coordination for you.
Sub-level 4 — Agent Teams (experimental): Sub-agents that communicate with each other + a supervisor agent. Enable via settings.json. Higher token cost due to coordination overhead. Best for: parallel features, cross-layer work, research, debugging.

Trap: Skipping the foundation

Without Levels 1–5 mastered, you get five agents producing five flavours of slop simultaneously. Every level feeds this one. There are no shortcuts. Diminishing returns also kick in past 2–3 parallel instances — more terminals does not equal more productivity.

Orchestration amplifies what you already have, for better or worse. If your context skills (Level 3), tool selection (Level 4), and skills (Level 5) are solid, orchestration multiplies them. If they're weak, orchestration multiplies that weakness.

The full journey

You went from commanding a tool → collaborating with a partner → curating context → equipping an expert → encoding your workflows → managing a team that manages itself.

Level 1: Command → Output
Level 2: Plan → Collaborate
Level 3: Curate → Context
Level 4: Equip → Tools
Level 5: Encode → Workflows
Level 6: Orchestrate → Scale

Next steps

Reaching Level 6 with a team requires Levels 1–5 to be solid. If you want to accelerate that journey and ensure the foundation is robust, we can design the complete training programme for your team.

Frequently asked questions

What is a git worktree and why is it important for orchestration?: A worktree is a working copy of the repository in a separate directory, linked to a different branch. It allows multiple Claude instances to work on the same project without interfering with each other — each in its own space, on its own branch. At the end, you merge the work. Without worktrees, parallel agents would overwrite each other's files.
How many parallel instances are optimal?: Diminishing returns generally kick in at 2–3 parallel instances. More than 3 simultaneous instances increases coordination cost (yours and the tokens) without proportional productivity gain. The exception is projects with genuinely independent workstreams — frontend, backend, tests — where 3–4 agents do produce clear gains.
How do I divide tasks between agents effectively?: The basic rule: each agent needs a clear objective, enough context to work autonomously, and access only to the files it needs. The best divisions are by layer (frontend/backend), by feature (auth/payments/notifications), or by phase (implement/test/document). Avoid divisions where agents need to edit the same files.
Is Agent Teams ready for production?: Agent Teams is experimental at the time of writing. It works well for research and development use cases where coordination overhead is acceptable. For production projects with deadlines, the recommendation is sub-level 2 (manually managed worktrees) until the feature matures.
Does Level 6 require advanced git knowledge?: It requires comfort with basic git concepts (branches, merge, conflicts) and the ability to resolve merge conflicts. It doesn't require being a git expert. The --worktree command is straightforward; the most complex part is usually managing the final merge when multiple agents have worked on adjacent areas.