Feedback on a starter kit for background agents (WINK / Python)

Hi HN — I’d love feedback on a starter repo for building unattended/background agents in Python:

https://github.com/weakincentives/starter

It’s built on top of WINK (Weak Incentives). The goal is to make the “agent definition” a first-class artifact: prompt structure + tools + policies + feedback/evals, separated from whatever runtime/harness drives the loop.

The demo is a toy “secret trivia” agent, but it showcases: prompt sections that bundle instructions + tools, progressive disclosure, tool policies (ordering constraints), feedback providers, and evaluators.

Questions: - How are you approaching building agents that run unattended? - Are these abstractions useful, or would you do this differently? - What works? What doesn't? What feels weird?

Thanks for any critique.