Claude Code multi-agent orchestration is the set of ways 📝Claude Code splits work across more than one Claude at a time: subagents inside a session, background sessions you dispatch, agent teams with a lead, scripted workflows, and projects run by a coordinator, with cross-session messaging connecting the sessions you run yourself. The choice comes down to three questions: who coordinates, whether the workers need to talk, and whether they touch the same files.
Why It Changed in 2026
📝Anthropic rebuilt this surface over four months. In early July, subagents began running in the background by default (v2.1.198). On August 7, v2.1.224 let separate sessions message each other. In mid-August, fork mode turned on by default (v2.1.232), so a subagent can inherit the whole conversation instead of starting cold. On September 17, Projects arrived, giving Claude a coordinator that runs a stream of tasks as parallel cloud threads. Each step raises the ceiling and the bill: every subagent, teammate, thread, and workflow agent spends its own tokens against the same plan limits, and Anthropic's own October 22 guidance urges folding subagents back into a single loop when the work does not genuinely split.
Delegating Inside One Session
📝Claude Code Subagents are the base primitive: workers that run a side task in their own context and return a summary, so logs and searches never flood the main conversation. Forks are subagents that inherit the full conversation and prompt cache, covered in 📝Claude Code Forked Sub-Agents — Quick Reference. When a job outgrows a handful of subagents, 📝Claude Code Dynamic Workflows move the plan into a script that can fan out to hundreds of agents and cross-check their findings. The general pattern behind all three is described in 📝Subagents.
Running Parallel Sessions
📝Claude Code Agent View opens with claude agents and keeps the user as dispatcher: each background session is a full conversation, moved into its own 📝Worktree before it edits files. Agent teams, documented here as 📝Claude Code Swarm Mode, go further: a lead session spawns teammates that share a task list and message each other directly. Teams are experimental, off by default, and share one checkout, so each teammate needs its own set of files. 📝Agent Swarms in Claude Code compares the spawn modes side by side.
Coordinators
A 📝Claude Coordinator scopes work, delegates it, tracks state, and assembles results without doing the work itself. In 📝Claude Code Projects, the project conversation plays that role: it routes each message to an answer, a new thread, or an existing thread, and an Overview pane shows which threads are ready for review or waiting on you. On the API side, 📝Claude Managed Agents configures coordination through a multiagent block with subagents, workflow runs, and an optional advisor model.
Sessions That Talk to Each Other
📝Claude Code Cross-Session Messaging connects sessions Claude did not spawn. One session finds another with ListAgents and sends it plain text with SendMessage: a breaking change, a settled decision, or a status report. Same-machine messages stay on the machine, and a message can never approve a permission prompt or change another session's configuration, which keeps parallel work coordinated without letting one session act through another.
Related
- 📝How to Start a Project in Claude Code — set up a coordinator and send it a first batch
- 📝How to Dispatch Subagents in Claude Code — delegate side tasks within one session
- 📝How to Message Between Claude Code Sessions — connect parallel sessions
- 📝Claude Code Multi-Agent — Cheat Sheet — every command, setting, and gotcha on one page
- 📝Multi-Agent Orchestration at Scale — production architecture beyond a single tool
- 📝AI Agent Swarms vs. Sequential Review: When Each Wins — when not to parallelize
- 📝Claude Code Features — the full Claude Code feature index