Notes
🔍

Multi-Agent Communication

Two agents on the same job that cannot talk to each other are just two agents doing the job twice. Multi-agent communication is the layer that stops that — how agents hand over work, announce results, and correct one another.

The shapes it takes
Systems pick from a small set of arrangements, and the choice decides who reads what:

  • Direct messages: one agent addresses another and passes the task along, the pattern behind a handoff.
  • Broadcast: everyone posts into a shared thread and everyone reads it — AutoGen's group chat.
  • Shared state: no messages at all; agents read and write a common store, as in LangGraph's graph state or a blackboard.
  • Cross-process protocols: when the other agent belongs to someone else, a wire standard such as Agent2Agent carries the exchange.

Messages are either free text or a structured envelope with sender, task id and a typed payload. The structured form is what lets a receiving agent route automatically instead of parsing prose.

Why it matters
Every message is tokens in somebody's context window, so a broadcast to five agents costs five times over and fills each context with talk rather than work. The failures are recognisable: a result nobody announced gets recomputed, a fact degrades as it is relayed and re-summarised, and agents spend rounds agreeing with each other while nothing gets verified. Deciding who hears what, in what shape, and when the conversation must stop is as much of the design as the agents themselves.

Multi-agent communication is how agents within one system exchange work: direct messages for a handoff, a broadcast thread everyone reads, shared state instead of messages, or a wire protocol such as Agent2Agent across process boundaries. It is what prevents duplicated effort, and it is not free — every message consumes context window in each agent that receives it, so who hears what is a real design decision.

Put four people on a project and the work is only half the problem; the rest is who tells whom what. The same is true when several AI assistants share a job.

Multi-agent communication is how they pass things between each other. Sometimes one hands the task straight to another. Sometimes they all post into a shared thread that everyone reads. Sometimes nobody sends anything at all and they simply write on a common whiteboard.

It matters because talking costs the assistants the same kind of effort as working. Too little of it and two of them do the same job twice; too much and they spend the whole day in a meeting.