Harnesses

A "harness" in the agent platform sits one level above the oracle loop: it runs many agent sessions on the same target, scores them against constraints, and converges toward a desired state. The agents package ships the harness protocol (HarnessProtocol.ts) and the data shapes (HarnessRun.ts), plus the AgencyInvokeTool that lets a harness spawn agency-profile sessions. Smithers (@open-matrix/smithers) is the application that drives convergence over GitHub issues today.

Status (2026-05-05). The shapes and protocol surface in @open-matrix/agents/src/harness/ are present and exported. Smithers consumes them. The wider workstream — the Stanford-style meta-harness with proposer agents, candidate evolution, and frontier promotion — is mostly TARGET STATE per WORKSTREAMS/stanford-harness/EVOLUTION-ARCHITECTURE-v3.md. Each page below is honest about what's wired vs roadmap.

Pages

• Coding harness — Smithers as the canonical coding harness.
• Evaluation harness — the evaluate profile and deterministic loops.
• Prompt and tool evolution — the candidate evolution surface (target state).
• Trace replay — replay activity frames against a frozen tool set.
• Frontier promotion — promoting candidates from staging to live (target state).

See also

• Reference: agent actor ops
• Tool testing

Source: projects/matrix-3/packages/agents/src/harness/, projects/matrix-3/packages/smithers/src/.