A Claude coordinator is the agent in 📝Anthropic's multi-agent products that scopes incoming work, delegates it to worker agents, tracks their state, and assembles the results, rather than doing the work itself.
In 📝Claude Code Projects, launched September 17, 2026, the project conversation is the coordinator. It decides whether each message gets an answer in place, a new thread, or a hand-off to an existing thread, and it sees what threads report back rather than every step they take. The coordinator has its own model and effort setting, separate from the threads, and Anthropic compares briefing it to briefing a chief of staff. Preferences given to it, such as "run at most two threads at a time", are saved to project memory and followed, but are not enforced caps.
In 📝Claude Managed Agents, coordination is configured on the agent through a multiagent block. The earlier coordinator type let an agent delegate to a listed roster; the current multiagent_20261001 type adds subagents with persistent threads, dynamic workflow runs, and an optional advisor model. Delegation is one level deep, a roster holds up to 20 unique agents, and a session runs at most 25 child threads, all sharing one sandbox and filesystem.
Two neighboring roles differ in who holds the plan. The lead in Claude Code agent teams (see 📝Claude Code Swarm Mode) supervises teammates that also message each other directly, while 📝Claude Code Dynamic Workflows move the plan out of any agent and into a script. In October 2026 guidance, Anthropic noted that fewer subagents with progressively disclosed context can raise quality while lowering cost and latency, so a coordinator earns its overhead only when work genuinely splits.
📝How to Start a Project in Claude Code shows how to brief a coordinator, and 📝Claude Code Multi-Agent Orchestration compares it with every other way to run Claude in parallel.