Claude Construction: Multi-Agent Coordination

An earthmoving site is a coordination problem. An excavator can't load a truck that isn't there, and a truck shouldn't move while a bucket is swinging over its bed. We gave each machine its own AI agent and let the agents work it out by talking to each other.

The setup

The value of an agent often stems more from its ability to collaborate with other agents then its own merits in isolation. This is particularly true in the physical world where a number of practical factors motivate the need for multiple agents to collaborate. In this experiment we explore coordination in the context of toy construction vehicles in a sandbox.

Each vehicle gets one agent. That agent owns the machine's perception (an overhead camera plus whatever the vehicle reports about itself), its action space (drive, steer, swing, boom, bucket, dump), and its own task state. There is no central planner holding a global schedule. The only shared structure is a message channel and a common coordinate frame.