The BoB Architecture Harness
A guided tour of the harness around the model: the four services, the fleet of roles, and why the model is the swappable part. A short video overview, plus a full podcast deep dive.
Most AI tools treat the model as the product. BoB treats it as a part, like a CPU you can swap. What makes the work reliable is the harness around the model, and this is a tour of it.
Watch the overview (about 7 minutes) for the shape of the system, then listen to the deep dive podcast below for the full walk through how the pieces fit.
What the harness is
- Four services.
coreroutes work, dispatches it across a fleet, verifies results, and keeps a durable record.gatewayis the door (web, API, login, approvals).pipelineis a thin planning wrapper.appis the desktop daily driver. - A fleet of roles, not vendors. Work is expressed as an apex that plans, workers that do the labor, and critics that adversarially check it. Any given role can be filled by whichever model is cheapest and best this week.
- A verification spine. Every meaningful step is checked by a different family of model before it’s trusted, so no single model is load-bearing and no single mistake goes unchallenged.
Why it’s built this way
When the model is swappable, nothing you build can be repriced, deprecated, or export-banned out from under you. You own the harness; you rent the intelligence. The video and podcast below unpack how that plays out in practice.
For the written version, see How BoB is built and the BoBClaw whitepaper.