Manifesto

Agentic computers already exist. They browse, write, file, reconcile and reply — around the clock, without complaint. The models will keep getting better without our help.

What does not exist is the layer that makes an office actually trust one:

• A queue everyone can see, instead of prompts lost in chat histories.
• Guardrails, so the agent asks before it spends, deletes or sends.
• A log, so "what did it do?" has an answer that holds up.
• A watchdog, so a stuck agent is restarted before anyone notices.

Nobody hires a person and never checks their work. Yet that is exactly how most teams run their agents today — full access, zero review, and hope as the safety model.

Harness takes the opposite bet. Treat the agent as a capable but junior colleague: give it real work, real deadlines and real working conditions — and put the supervision in software, where it is cheap, instead of in people, where it is expensive.

That is the whole product. Not a smarter agent. A dependable one.