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

Two kinds of graph.

The domain graph is the set of links that already exist in your data: accounts to devices, services to deploys, suppliers to SKUs, facts to facts. The control graph says which agent goes next and which one is responsible for the reply.

domain    world already linked
          account · service · supplier · fact

control   who goes next
          who owns the reply

Most of these jobs use the first kind. One agent following the links is often enough, as Part 3 explained.

Fraud investigation and entity resolution

A hidden fraud ring does not appear in a single row of data. The links between account, device, and transaction are what reveal it. Domain graph.

account ── device
    │         │
    └── txn ──┘

You follow the links and record what you find. You can check several sources in parallel, but the merge still has to see every result (Part 2). Having five columns in a spreadsheet is not a reason to create five agents.

IT incident response and root-cause analysis

A cloud error is not a root cause until it matches a deploy or a tenant change. Domain graph.

error → service → deploy → tenant
              missing write = lying edge

Walk the dependencies. If the deploy step recorded nothing, the next step makes up a tidy explanation. The problem is a connection passing on missing information, not a weak prompt. Part 1 drew the map, and Memory covers how results are saved.

Supply-chain impact analysis

One supplier delay fans out through plant, SKU, and lane. Domain graph.

supplier delay
    → plant
    → SKU
    → lane

The merge is the impact picture: who saw every downstream write? You do not need a team of logistics agents to trace that chain of effects. You need the edges, and a stop when the picture is good enough.

Advanced research and multi-step questions

A hard question breaks into several steps. Follow the exact relationships between facts. Loop through tools until a check passes. Domain graph plus one agent loop.

question → walk → tool → check
                ↑          │
                └── loop ──┘
                     stop

Prompts and a stop still matter. This is not five researchers, and it is not a comparison of products. The method is: break the question down, follow the links, and check the answer.

Multi-agent workflow orchestration

Planner, coder, reviewer. Control graph.

planner → coder ⟲ retry
              ↓
          reviewer → reply

What’s coming next already showed them in a line. Real work branches, retries, and passes tasks between agents; it does not move in a fixed line. The connections are still written down. Orchestration decides when each step runs. Part 1 named the nodes. Part 2 named the merge. The agents do not decide for themselves how they are organised.

Takeaway: The world is already a graph. Add extra nodes only when the instructions, tools, or rules change.