Multi-Agent in Production
Four ways to run several agents, the named forces that justify a separate deployment, and why the default is one process.
Last updated
After this section you can
- Choose one process or separate deployments, and name the force behind the choice
- Work out the latency and availability of a chain of agents, then set its deadlines
- Pick a wiring for a workload, and name what that choice makes you own
33
Multi-Agent in Production
Four ways to wire the same four agents, the named force that justifies a network hop, and why the default is one process.
Key idea
Agent-to-agent calls over a network are slow, costly and hard to verify, so every separate deployment has to earn its hop with a named force. Start with one agent in one process. Once a call leaves that process, contracts, budgets and traces are the design.
Same four agents, four wirings: each wire you add is a failure mode
Amber: the agents. Purple: state the wiring keeps. Teal: the wires, the subject of this lesson. Pans sideways on a phone.