Epilogue

A book like this is a snapshot of an architecture discipline that is still being written. The cognitive patterns will move; the models will improve; the Skills standard will evolve; new failure modes will be discovered and named. What is more durable, and what this book has tried to develop, is the shape of the discipline: probabilistic reasoning components, bounded by deterministic infrastructure that the surrounding system controls; governance enforced as structure rather than convention; memory architected rather than accumulated; failure modes cataloged rather than discovered at incident time; traces kept as the system of record.

If the book has done its job, the reader closes it with three things: a vocabulary for talking about agentic-system architecture with peers, a set of templates (the bounding YAML, the governance pipeline, the trace schema, the skill manifest, the harness loop, the Concord worked example) to adapt, and the conviction that the architecture around the model carries more of the system’s reliability than the model itself does.

Build the architecture around the agent, not under it. The rest follows.