Your army of agents needs contracts

I watched a multi-agent pipeline give silently wrong answers for 6 hours, with no errors and no crashes.

The cause was one team changing an agent's output format without the downstream team knowing.

A contract states the output shape, the types (Pydantic/JSON Schema) and the fallback sent on failure.

Applying contracts at agent boundaries is the cheapest reliability investment I've found.

An I/O contract states the output shape and types.

A degraded-mode contract adds the failure path, meaning what gets sent when full data isn't available.

A2A Protocol specification

Pydantic docs