![# Harness principles We are learning how to make AI a dependable participant in ongoing work by building harnesses and observing their results. Through this work, we value: - [ ] Shared task state, visible decisions, and clear ownership over isolated conversations. - [ ] Business purpose, relevant context, and technical understanding over generic intelligence alone. - [ ] Useful initiative and follow-through to resolution over activity and notifications. - [ ] Explicit authority and enforced boundaries over assumed permission. - [ ] Verified outcomes and recoverable actions over convincing answers and claims of completion. - [ ] Corrections that retire obsolete work and improve behavior over history that merely accumulates. - [ ] Bounded delegation and verifiable handoffs between agents over manual coordination between products. We judge a harness by the useful work it completes and the human attention it frees. It should know when to act, when to involve people, and when to stop.](https://pbs.twimg.com/media/HWHlSl5WUAA0VDj.jpg?name=small&format=webp)
ALT # Harness principles We are learning how to make AI a dependable participant in ongoing work by building harnesses and observing their results. Through this work, we value: - [ ] Shared task state, visible decisions, and clear ownership over isolated conversations. - [ ] Business purpose, relevant context, and technical understanding over generic intelligence alone. - [ ] Useful initiative and follow-through to resolution over activity and notifications. - [ ] Explicit authority and enforced boundaries over assumed permission. - [ ] Verified outcomes and recoverable actions over convincing answers and claims of completion. - [ ] Corrections that retire obsolete work and improve behavior over history that merely accumulates. - [ ] Bounded delegation and verifiable handoffs between agents over manual coordination between products. We judge a harness by the useful work it completes and the human attention it frees. It should know when to act, when to involve people, and when to stop.








