Benchmark the harness, not just the model
The HarnessRouter infrastructure layer transforms existing harnesses, such as Codex and Claude Code, into plug-and-play agent backends. It enables products to run tasks, retrieve results, and switch harnesses through a unified API. The Unified Harness Protocol (UHP) powers HarnessRouter, providing an OpenAI Responses-compatible API that manages persistent sessions, streaming progress, files, artifacts, cancellation, and structured failures.
Eight harness models can be configured on a single task, with each configuration offering a best and worst result set. The lowest-cost and fastest configurations differ based on the task. The methodology utilizes 31 harness bases, including Codex, Claude Code, Hermes, Pi, DeepSeek Harness, OpenCode, and various other models. The support matrix details which harnesses run on which provider.
To get started, users can run the product locally, integrate it into their product, utilize starter kits, or access managed Cloud and self-hosted Community Edition options. The process involves pulling the latest Docker image, running it with Docker, and connecting a model provider. Users are not required to have a HarnessRouter account or bundled model/trial keys.
The Console and Gateway run unprivileged, while agent processes operate under the session's user. Upon the first launch, the container keeps running, and users can access the Console at http://localhost:3000 or a chosen host port. After signing in, users can add their own provider key to authorize model requests. The agent harness and model can be set to Auto or chosen manually, and users can save the setup.
The agent harness remains fixed once a setup is saved, and users can add plugins, skills, MCP servers, and environments for the entire workspace under Add-ons. Running product tasks with built-in or custom harnesses as pluggable agent backends is possible, with tasks and their transcripts appearing in the Console workspace. To enable server-sent events, set stream: true and send W3C Trace Context headers for traceability.
For idempotency, use a unique Idempotency-Key for each new logical request, and never include sensitive information like tokens in the prompt.
Written by urgent.news from Daily Dose of DS's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
