An AI agent that hires physical-world work today writes a separate integration for every worker provider it talks to. Software-tool protocols do not help: a tool cannot drop a payload halfway across a room. The Worker Context Protocol (WCP) is one wire contract for an agent to dispatch work to a human, an autonomous robot, a teleoperated system, or a hybrid of these. The matching engine and the attestation verifier never read the worker’s class. They read the kind of evidence the worker emits. One code path therefore accepts a human’s phone GPS track and a robot’s onboard odometry as proof of the same task. WCP is ten remote-procedure calls (eight task-lifecycle plus two administrative), a typed attestation primitive with an M-of-N evaluator, an attestation retry mechanism for evidence the verifier rejects, and a hashlinked audit chain. A Python reference implementation ships with TypeScript, Rust, and Go client SDKs. Seven worked agents across institutionally distinct domains and a working two-coordinator federation demo exercise the surface end to end; a three-level conformance suite gates the claim. WCP coordinates and records. Real-time control, safety certification, swarm orchestration, and accounting live elsewhere in the stack.
Ambar (2026) studied this question.