Field Notes

What I actually care about in agent systems.

This is the short version I want people to see when they land from X: practical routing, visible failures, and local models that earn their place.

Routing

A harness should make the local-vs-cloud choice boring and explicit. If it hides the decision, the system is already harder to trust.

Failure visibility

I care more about replayable failure paths than polished demos. If I can’t see what changed, I can’t tell whether the system got better.

Local inference

Local models are useful when the loop is fast enough to matter and the setup is honest about tradeoffs.