Graphsmith
An AI-native planning and execution system built around typed graph programs.
Language models are good at semantic intent but unreliable at emitting large executable structures directly. Graphsmith separates these jobs: a model proposes a typed intermediate plan, while deterministic machinery compiles, validates, repairs, executes and traces the graph.
Current direction
The project explores guarded execution, loop lowering, local repair, reusable skills, trace-derived promotion signals, and trust-aware remote reuse. Its central bet is that inspectable graph structure gives models—and people—a better substrate for understanding and repairing programs than opaque one-shot generation.