Two jobs
- Your product
- Your product: UI, login, tenants, permissions, business data, server routes, storage, previews, and downloads.
- Runtime agent
- The configured HarnessRouter agent that receives one authorized user task and returns progress, text, files, and artifacts.
Top questions
- Does HarnessRouter replace my app UI?
- No. Your app keeps the interface, brand, login, permissions, customer context, previews, downloads, and follow-up flow.
- Where do Cursor, Codex, or Claude Code fit?
- They are build-time coding agents. Paste AGENTS.md into them so they can add the HarnessRouter integration to your product.
- What does HarnessRouter receive at runtime?
- One product task from an authorized user, plus the files and context your server decided that task may access.
- How do I choose harness and model?
- Start with candidate configured agents, run representative tasks, then compare Trace, stream behavior, recovery, files, artifact quality, cost, and latency.
- Who authorizes sessions and files?
- Your server does. Treat HarnessRouter session and file IDs as opaque backend resources mapped to your own user or tenant records.
Example
Keep build-time and runtime separate
Ask your coding agent to build: A contract review product with upload, progress, preview, download, and revise. What your product sends to HarnessRouter at runtime: Review this uploaded agreement and return a checklist, redline notes, and generated files.