Why I started Truss Labs
Notes on incorporating in New York, picking a C-Corp, and why I'm working on agent infrastructure instead of another agent product.
What I'm building
Truss Labs is a one-person R&D lab working on the layer underneath AI agents — state, audit trails, and control interfaces. The "brain" (the model) has improved fast; the scaffolding around it hasn't.
Most agent products today run reasoning inside a linear chat box, log nothing you can replay, and have no clean way to intervene mid-run. I think that's the real gap worth working on.
I. Soul Kernel
A local state store for agent memory — sessions, tasks, and traces on disk, not inside a vendor's database. Runs as a shell harness so it composes with existing tools.
II. Control Plane
A UI for reading and steering agent runs beyond the chat box. Render reasoning as a tree you can branch, inspect, and replay. Early prototype.
III. Primitives
Small CLI tools (soul-query, soul-trap, soul-ingest) for querying traces, halting on confidence drops, and loading context. Unix-style, piped together.
Status: pre-product. I'm writing about the work as I go rather than waiting for a launch.