Adding more AI personas does not make a system.

What’s coming next already named the linear chain: researcher → writer → reviewer. Orchestration decides when each step runs. Memory is where results are saved. All three assume one idea, and this series explains it: a multi-agent system is a graph.

A node is a specialist agent with its own loop. It still needs a job description and a point where it stops. An edge says which agent goes next and which one is responsible for the reply. State is the information the next node is allowed to see. A single loop is the smallest possible graph. A system becomes multi-agent when you need more than one node and you write down how they connect.

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

The four parts draw that map, then walk it:

Part 1  path          one walk
Part 2  fan-out       then merge
Part 3  one node      split only if the contract changes
Part 4  jobs          domain graph vs control graph

One survey calls this design work graph engineering. This series stays focused on multi-agent systems and uses the graph as a way to draw them.

The series

  1. The map — nodes, edges, state, stop; a chain is only one pattern
  2. Fan-out, then merge — who saw both writes; one sourced example
  3. When not to split — start with one agent; agreement is not proof
  4. Where it shows up — five jobs; domain graph vs control graph

Sources