Question
What is the smallest useful runtime abstraction for reliable AI agents?
Hypothesis
An agent runtime needs typed events, tool adapters, policy guards, and replay before it needs a large framework.
Method
Sketch the event model, then prototype a run loop with tool calls, retries, and resumable state.
Prototype
A local TypeScript runtime that stores plan, action, observation, validation, and handoff events.
Notes
The runtime should make the default failure mode visible rather than magical.
Results / Open Questions
Open question: which events belong in durable state, and which can be reconstructed from traces?
References
Placeholder for durable execution, workflow engines, and agent orchestration notes.