Part 1 of 4 — Multi-agent systems are a graph

Adding agent names does not give you a map.

The introduction said a multi-agent system is a graph. This part only defines its parts. What’s coming next already showed the linear chain: researcher → writer → reviewer. We will not repeat it here. A chain is only one pattern. Every map has the same four parts.

A node is a specialist agent with its own loop. It still needs a job description and a point where it stops. Orchestration covers how that loop is run, so this page does not.

An edge says which agent goes next and which one is responsible for the reply. If control passes to the next agent but the results do not, the next agent cannot see what happened. It is only there for show.

State is the information the next node is allowed to see. Memory covers how results are saved. If nothing is saved, the next call starts with no memory of earlier steps.

Each node also needs a stop. A specialist with no limit is not independent. It is a branch that can run out of control.

node   = specialist + its loop + its stop
edge   = who goes next / who owns the reply
state  = what the next node may see

Part 10’s researcher → writer → reviewer is a path along those edges:

researcher → writer → reviewer → reply
              one walk

This is useful, but incomplete. Work that needs two specialists at the same time cannot follow a single path. Part 2 is fan-out, then merge.

Prompts and loops are still needed. Each node still needs a job description, and each branch still needs a stop.

Takeaway: Draw nodes, edges, state, and a stop. A chain of personas is only one walk across that map.