AI agent orchestrators: how they work and which to use for coding

An agent orchestrator is the agent or system that coordinates other agents. It breaks a goal into tasks, starts a worker for each task with a focused prompt, keeps the workers from getting in each other's way, and decides what happens with the results: combine them, verify them, retry, or ask a person.

Frameworks such as LangGraph, CrewAI, and the OpenAI Agents SDK let developers build an orchestrator into their own product. This guide covers orchestrators for coding agents, which come in two forms. Checked October 7, 2026 against vendor documentation. Replicas publishes this guide and is one of the platforms below.

An orchestrator agent

A coding agent session that plans and delegates. The main Claude Code or Codex session spawns subagents, each with its own context window, and merges their findings into one answer.

An orchestrator platform

Software that takes tasks from a queue, prepares an environment for each one, runs a coding agent in it, and routes the result back for review. The orchestration happens across sessions and machines rather than inside one conversation.

Almost every orchestrator follows the same loop. What changes between products is where the workers run and how isolated they are.

Plan

Split the goal into subtasks that can proceed without waiting on each other.

Dispatch

Start a worker per subtask with its own prompt, tools, and model.

Isolate

Give each worker its own context window and, for code, its own checkout or machine so edits do not collide.

Collect and verify

Bring results back, run tests, and review the changes with something other than the worker that made them.

Decide

Merge, retry, reassign, or hand the decision to a person.

• Every worker is a separate model session, so orchestration uses more tokens than one agent doing the same work.
• Parallel workers only help when subtasks are independent. Sequential steps and edits to the same files are usually faster in a single session.

The main Claude Code session is an orchestrator. It delegates to subagents, each running in its own context window and returning a summary, so exploration and logs stay out of the main conversation. Claude Code ships with Explore, Plan, and general-purpose subagents, and you can define your own as Markdown files with YAML frontmatter in .claude/agents/ for a project or ~/.claude/agents/ for yourself.

By default up to 20 subagents run at once and they can nest three levels deep. Adding isolation: worktree to a subagent's frontmatter gives it a temporary git worktree, so parallel edits do not touch the same files. The /batch command uses this to fan one change out to many worktree subagents that each open a pull request.

Agent teams, which are experimental, take this further. With CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS set to 1, the lead session spawns teammates that are full Claude Code instances, share a task list, and message each other directly.

Codex also runs subagent workflows, enabled by default in current releases. It delegates when you ask, or when AGENTS.md or a skill tells it to, then spawns agents in parallel and collects their results into one response. That suits codebase exploration and multi-step feature plans. For example, ask it to "spawn one agent per point" or "delegate this work in parallel".

Codex ships with default, worker, and explorer agents. Custom agents are TOML files with a name, a description, and developer_instructions, stored in ~/.codex/agents/ or a project's .codex/agents/, and can set their own model and sandbox mode. Subagents inherit the current sandbox policy and approval settings, and each one uses its own tokens.

When the work is a team backlog rather than one conversation, the orchestrator moves into a platform. These differ in which agents they run, where those agents run, and what starts the work. The agent orchestration platforms guide compares nine of them in detail.

Pick the lightest orchestrator that covers the work. Most teams end up using more than one: a platform hands out tasks, and the agent inside each task still delegates to its own subagents.

One task with research or review on the side

Subagents in Claude Code or Codex. No setup beyond an optional agent definition.

One problem that benefits from debate

Claude Code agent teams, such as several teammates testing competing hypotheses for a bug.

A few independent tasks you steer yourself

Separate sessions in git worktrees on your own machine.

Tasks you want off your laptop, with one vendor

That vendor's cloud: Claude Code cloud sessions, Codex cloud, or Cursor Cloud Agents.

A team backlog across agents and tools

An orchestrator platform such as Replicas, where tasks start from Slack, Linear, GitHub, or a schedule and each runs the agent you choose.

Replicas includes a 14-day free trial with no credit card required. Connect a repository, define the environment once, and assign tasks from the dashboard, Slack, Linear, or GitHub. Each task runs in its own Linux VM with Claude Code, Codex, Cursor, or OpenCode, and comes back as a pull request your team can review.

• How to orchestrate coding agents

  Five ways to run several agents, from subagents to cloud VMs, with setup steps.

• Agent orchestration platforms

  Patterns and nine orchestration platforms compared.

• Multi-agent orchestration for coding agents

  Queues, isolation, review, and merge strategy for a fleet of agents.

• Claude Code subagents

  Anthropic's reference for subagent definitions, delegation, and limits.

• Claude Code agent teams

  Anthropic's reference for experimental agent teams.

• Codex subagents

  OpenAI's reference for built-in and custom Codex agents.