Beyond Subagents: Claude Code's New Agent Teams Let AI Workers Message Each Other Directly
The standard multi-agent pattern in most frameworks follows a hub-and-spoke model: a main agent spawns subagents, those subagents do work and report back, and the main agent synthesizes results. It works, but the bottleneck compounds as parallelism increases — every inter-agent communication routes through the coordinator, creating latency and a single point of failure for coordination state. A new experimental feature in Claude Code takes a different approach.
Agent Teams introduces peer-to-peer direct messaging between independent Claude Code sessions. A team lead still spawns teammates and each has its own context window, but coordination no longer flows through the lead. Workers claim tasks from a shared list autonomously, and when one agent needs output from another, it can message that peer directly rather than escalating to the coordinator. Developers also gain the ability to address individual teammates directly — a meaningful improvement for debugging specific parallel workstreams without interrupting the rest of the team.
The guide draws a practical distinction about when this architecture justifies the added coordination overhead. The signal is genuine parallel workstreams with inter-stream dependencies: tasks that can run concurrently but need to exchange results mid-execution. Where that pattern doesn't exist — isolated parallel tasks with no shared state — the simpler subagent model remains the right call. The guide includes a direct comparison table with the older builder-validator chain pattern and concrete prompting patterns for structuring the shared task list.
The shift from hub-and-spoke to peer-to-peer is a structural change in how multi-agent coding systems can be designed, and this is the first detailed public walkthrough of the pattern as it ships in Claude Code today.