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.
