Where does static analysis end and runtime evidence begin for AI agents?

[hpy1968]

The "AI-native era" framing resonated. I'm working on something adjacent: Retrace, an open-source reverse debugger and record-replay tool for Python (github.com/retracesoftware/retracesoftware). The focus is the runtime side of the same agentic workflow. When static checks pass but a test, CI job, or production execution still fails, Retrace captures the actual failed execution so a developer or AI agent can inspect what really happened, including stepping backwards through it.
The complementarity as I see it:

• Pyrefly gives the agent fast static feedback: types, structure, imports, IDE context.
• Retrace gives the agent runtime evidence: the failing execution path, variable state, replayable context.

This feels especially relevant for AI-generated Python, where the agent can fix type errors quickly but still struggles with runtime or data-dependent failures from logs and source alone.
How does the Pyrefly team think about the boundary between static and runtime feedback in agentic workflows? Pre-run / CI verification, or part of a broader loop where agents combine static checks with runtime traces and debugger state? Congrats on the launch either way.