I've been researching sandboxing (bwrap et al) focusing on good UX for system tests to empower non tech users in a deeply technical space. They've been coding. They're writing agent skills. No evals yet. This space is critical but UX is too. Are you targeting that type of user as a design principle?
published: No news is good news
alexhans.github.io/posts/series...
About building systems that don’t demand constant attention: fewer dashboards, fewer alerts, clearer expectations. If things are working, you shouldn’t have to look. On feedback loops and boring reliability for small teams.
published: Error compounding in GenAI systems alexhans.github.io/posts/series...
Small mistakes don’t stay small once you chain agents. This is why evals + explicit expectations matter. Written for anyone building multi-step workflows. Many find the same learnings independently.
published: Building agent skills incrementally
alexhans.github.io/posts/series...
Not about frameworks or “AI platforms” - about treating agents like software: small skills, explicit expectations, evals, repeat.
Written for non-traditional builders (e.g. economists). you need feedback loops.
Hey Simon Willison - this resonates. I’m exploring how evals help non-tech folks ship end-to-end. ai-evals.io isn’t a product (no selling) - just a community site. Are you using evals? Seen them work with non-engineers? Any advice on advocating the mindset shift?
