WINK (Weak Incentives): portable agent definitions for unattended agents (Python)

Hi all - sharing an Apache-2.0 Python framework for unattended/background agents.

Core idea:

  • Separate “agent definition” (prompt/tools/policies/feedback you version/test) from the runtime harness (loop/sandbox/orchestration).
  • Prompt is a hierarchical doc; sections bundle instructions + tools; sections can be enabled/disabled.

Notable features:

  • Transactional tool calls + sandboxed VFS/mounts
  • Policies as invariants over tool usage
  • Debug UI / state snapshots
  • Adapters for different harnesses

Starter (minimal runnable example): GitHub - weakincentives/starter: Starter kit for Weak Incentives (WINK)

Core lib: GitHub - weakincentives/weakincentives: Tools for developing and optimizing background agents.

Feedback I’d love:

  • What “unattended agent” workflows are you trying to productionize?
  • Does prompt-as-agent (sections+tools co-located) feel like it would help or get in the way?
1 Like