Benjamin Borowski 🍁 @typeoneerror.com · 30/09/2026i don't suppose harnesses can be done by definition, both by the possibility that tool calls aren't guaranteed by context and predictably at some point you're going to change models. curious what your testing harness knows. 220
Kai 🌟 @kaidav.is · 30/09/2026i agree. i don't think deterministic returns are possible. my goal is to surface *some* of the relevant contextual data for the agent at the right time + place and give pointers for an appearance of continuity. reduce the cold start new pane 'you don't know the thing we worked on for a week?' issue. 020
