Specialised AI roles
A small set of agents, each with a clear job. Routing is deterministic; coordination is explicit.
What this layer is
Above the metadata layer sits a small roster of specialised agents — one for analysis, one for testing, one for build, one for governance, one for capturing what was learned. Each has a written charter.
Each role has its own concerns and its own constraints. An analysis agent is not allowed to write production code; a build agent is not allowed to invent business rules. The division of labour is the safety mechanism.
The most autonomous of these roles is the Epic Loop: a deterministic controller that drives an AI “maker” to build a program’s full, Docker-validated test suite end to end. The controller is code, not judgement — it decides what runs and in what order; the AI does the work inside those rails. That pairing is what lets the platform take on real delivery without giving an unconstrained model the keys.
Failure mode it prevents
Tribal negotiation over who owns which task; one generic assistant trying to be everything at once.
A focused walk through the methodology.
A working session covering all five layers and how they would map to the systems you actually have to modernise.