agent-core is an in-progress project: a proper agent runtime, rather than the pile of one-off scripts that tend to grow around this kind of work.

The shape of it

It models a task as a graph. Work flows through nodes, each doing one well-defined thing before handing off to the next. A run is a sequence of small, inspectable steps — so when something goes wrong, it is clear where.

Underneath that are a few deliberate separations: the model sits behind a provider interface, so swapping or combining models doesn't touch the graph; the agent acts through a constrained set of tools rather than raw access; and file changes pass through validation phases before they land.

Where it is going

This is a notebook entry, not a launch. The current focus is making runs fully resumable and tightening the validators that gate changes. The aim is simple: the next thing I build should sit on top of agent-core instead of reinventing the orchestration again.

More notes as it develops.