Agentic Assistant Architecture

In the context of specification-driven development, this document describes how a set of related role-scoped agentic assistants is structured. Each is a tool scoped to one human capability in the software development lifecycle: requirements, technical design, quality, project management, and the engineering of the assistants themselves. Each of these tools is based on three key concepts. Practice is how the work is done. Knowledge is what applying the practice produces, which is the deliverables themselves and the terminology that settles across them. Improvement is the other thing applying it produces, which is the record of where the practice did not fit. Applying practice produces knowledge and improvement; applying improvement produces practice. Professional practice ships inside the package, which is what makes it portable. Organisational practice, the deliverables and the improvement record live in an external data storage outside the assistant platform, in documents people author and own. A placement test decides which side of that boundary anything falls on, seven components implement the structure, and one project-management assistant published under an open licence is worked through them. Because each assistant follows one named governing document at a declared version, what it produced can be reconstructed afterwards, which is the property the narrow scoping exists to provide. A first generation of these assistants was built and deployed at a supply-chain software company and is described here as well, because the second generation's structure follows from three specific things the first could not do. Nothing in this document measures whether the structure works.