What is agent orchestration?
Agent orchestration is coordinating several coding agents on one body of work: split it, run the pieces in parallel or in sequence, check each, then merge. No single person coined the name; Anthropic described an "orchestrator-worker pattern" in June 2025. Orchestration multiplies whatever review discipline the human has: with a judge, tests and a merge gate it is a factory, and without them it is vibe coding at fleet scale.
Origin No single person coined the name, and orchestration is an old systems word. Anthropic described an orchestrator-worker pattern for a research system in “How we built our multi-agent research system” (opens in a new tab) on 13 June 2025, and Steve Yegge introduced Gas Town, his “fourth complete, functioning orchestrator of 2025” for coding agents, in “Welcome to Gas Town” (opens in a new tab) on 1 January 2026.
- In Polish
- Orkiestracja agentów
- Closest ladder level
- All levels, free guide
- Also searched as
- multi-agent orchestration, multi-agent development, multi-agent coding
Then $19.99 a month or $99.99 a year. Cancel from your account page. 30-day refund on a first purchase.
Why it matters.
Orchestration multiplies whatever review discipline you already have. Steve Yegge says of Gas Town, his orchestrator: "It’s also 100% vibe coded. I’ve never seen the code, and I never care to, which might give you pause." Cost scales with the number of agents: Anthropic measured "about 15× more tokens than chats" for its multi-agent research system and cautioned that "most coding tasks involve fewer truly parallelizable tasks than research". Parallelism pays only where each piece can be verified on its own.
Agent orchestration vs vibe coding
Orchestration can be done vibe-style or with verification, and the difference is what checks each unit before the merge.
| Question | Vibe coding | Agent orchestration |
|---|---|---|
| What it names | A way of working: accept what the model writes without reading the diff | A technique and a tool category: split the work, run several agents, merge the results |
| Who writes the code | The model | Worker agents, one per unit of work (a subtask, a lane or a worktree) |
| Who checks it | Nobody reads the diff; the human watches the behaviour | In the planner, workers, judge pattern: a judge agent and deterministic checks, then the human at merge |
| What stops the loop | The app appears to work | Each unit’s acceptance command, then the orchestrator’s merge |
| Where the name came from | Andrej Karpathy, post on X, 2 February 2025 | No single coiner; Anthropic’s orchestrator-worker pattern, 13 June 2025 |
| Where it breaks | Beyond throwaway work, nothing replaces the reader | Without checks it is vibe coding at fleet scale: Steve Yegge calls Gas Town “100% vibe coded” |
Planner, worker and judge
Every orchestration pattern wires three roles differently, and most failures come from one agent silently playing two.
| Role | Does | Must not |
|---|---|---|
| Planner | Splits the goal into units, names the files each unit owns, writes the shared contracts and states each unit’s acceptance check | Write implementation code |
| Worker | Implements or investigates exactly one unit, in its own worktree or VM | Touch files outside its unit, or edit the tests that judge it |
| Judge | Runs the deterministic checks first, then compares or grades what the workers produced | Edit the code it judges, or know which agent or model wrote which candidate |
Four orchestration patterns in the three tools
All three tools run all four patterns, with different building blocks: scriptable workflows in Claude Code, subagents plus cloud attempts in Codex, subagents and per-VM cloud agents in Cursor.
| Pattern | Claude Code | Codex | Cursor |
|---|---|---|---|
| Fan-out/fan-in | Subagents; /batch for 5 to 30 worktree units; dynamic workflows for dozens to hundreds of agents | Subagents capped by [agents] max_threads; codex --worktree per lane | Subagents, which can run on their own virtual machines since 2026-08-19; worktrees for your own lanes |
| Pipeline | Workflow phase() blocks with a JSON schema on each agent() call | Chained codex exec --output-schema FILE -o out.json calls | Automations start cloud agents on a schedule or on events |
| Best-of-N with a judge | A workflow that runs N candidates, then a judge agent() | codex cloud exec --env ENV_ID --attempts 3 (cloud is experimental) | The same prompt on several Cloud Agents, then a judge |
| Writer/reviewer | A read-only reviewer subagent; /code-review | codex exec review --base main or /review | Bugbot reviews pull requests as a separate agent |
How each tool does it.
Cursor, Claude Code and Codex each name this their own way. Each note carries the date it was checked.
Cursor
Subagents, which since the 2026-08-19 release "can now run on their own virtual machines"; Automations start cloud agents on a schedule or on events.
Orchestration patterns guide (Cursor docs), 2026-08-28
Claude Code
Subagents in .claude/agents/; /batch splits a change into 5 to 30 worktree subagents; dynamic workflows fan out dozens to hundreds of agents from a script.
Defaults cap 20 running subagents per session and 16 concurrent workflow agents, and one workflow run takes up to 1,000 agents. Agent teams are experimental and off by default.
Orchestration patterns guide, Claude Code 2.1.283, 2026-09-26
Codex
Subagents on by default, capped with [agents] max_threads; for code lanes, codex --worktree per lane or git worktree add plus codex exec -C <dir>.
Tokens spent by nested subagents count toward [goals] max_goal_token_budget. codex cloud exec --attempts takes 1 to 4 attempts at one task, and cloud is experimental.
Orchestration patterns guide, Codex 0.157.1, 2026-09-26
Questions about agent orchestration.
How is agent orchestration different from vibe coding?
They are not opposites. Steve Yegge calls his orchestrator Gas Town "100% vibe coded" and says he has "never seen the code" (1 January 2026), so orchestration can be done vibe-style. With a judge, deterministic tests and a merge gate, the same pattern is a software factory. Our reading: orchestration multiplies whatever review discipline the human has, and verification is what differs.
Is agent orchestration the same as agentic workflows?
No. One agent iterating on its own output is an agentic workflow; two or more agents coordinated by a lead agent or a script is orchestration. Anthropic’s definitions give the steps: a workflow, where code paths are predefined; an agent, which directs its own process; and a multi-agent system, where "a lead agent coordinates the process". Andrew Ng’s fourth pattern, multi-agent collaboration, is where the two overlap.
Is it worth orchestrating several agents for coding?
Only where the work splits into pieces that can be checked on their own. Anthropic’s multi-agent research system beat single-agent Claude Opus 4 by 90.2% on its internal research eval, which is not a coding result, and the same post says most coding tasks involve fewer truly parallelizable tasks than research. DoltHub’s Tim Sehn reported about $100 in Claude tokens for one hour of Gas Town (one user, one setup).
Is running agents in several worktrees the same as orchestration?
No. Running three agents in three worktrees is parallelism. Orchestration is the planning, assignment and merge around it: a plan that names the files each unit owns, a check for each unit and a merge order. This site’s Codex guide describes lanes that share no files, one agent per lane, merged in dependency order after each passes its own gate.
The vocabulary of AI-driven development
Every name below is defined against vibe coding: who writes the code, who checks it, and what stops the loop.
Read the long-form guide: agentic engineering vs vibe coding
Sources.
The primary sources outside this site that this page relies on.
- How we built our multi-agent research system (opens in a new tab) Jeremy Hadfield, Barry Zhang, Kenneth Lien, Florian Scholz, Jeremy Fox and Daniel Ford, Anthropic Engineering
- Welcome to Gas Town (opens in a new tab) Steve Yegge, Medium
- Agentic Design Patterns Part 5: Multi-Agent Collaboration (opens in a new tab) Andrew Ng, The Batch
- A Day in Gas Town (opens in a new tab) Tim Sehn, DoltHub
Keep reading.
The guides that go deeper, and the terms and comparisons next to this one.
In the docs
- The full A-Z glossarySubscription
- Vibe coding vs agentic engineering: 21 terms comparedFree
- Level 5: You Run the Software FactoryFree
- Multi-agent orchestration patterns: planner, workers, judgeSubscription
- Dynamic workflows and ultracodeSubscription
- Codex multi-agent workflows: parallel lanes, subagents and cloud fan-outSubscription
- Running 10 Agents at Once with herdr, Conductor and WorktreesSubscription
Related terms
Comparisons
Read the guides in the same words.
Open every guide with the 7-day free trial. Each term is defined once here, and the guides use it the same way.
Then $19.99 a month or $99.99 a year. Cancel from your account page. 30-day refund on a first purchase.