Deliberate Omissions

Things an agent roadmap usually includes that were left out on purpose:

Omitted	Reason
A multi-agent architecture	Genuinely useful when work fans out across many sources. It is also the most common way a project acquires four times the cost and no more capability. One agent that can prove what it refuses beats five that cannot. Add a second when a measurement says one is the bottleneck.
An agent framework	They give you the loop, which the SDK’s tool runner already gives you, and they hide the per-turn hook where authority belongs. The hard parts of this roadmap — authority, idempotency, compensation, evaluation — are yours in every framework.
A managed agent platform	Step 2 covers when it is the right answer, and it often is. It removes the loop and the hosting; everything this material is about remains. Use one on a paid project once you can build without it.
Retrieval over the client’s documents	A different problem with a different failure shape, covered by its own material. An agent that answers from documents and an agent that changes the world need different proofs.
Fine-tuning	Almost never the answer for tool use. The tool surface and its descriptions are where behaviour is shaped, they change in minutes, and they are inspectable.
Long-running agents and schedules	A real direction — an agent that watches a queue overnight. It multiplies the authority problem before you have solved the synchronous one. Add it after step 12 exists.
Voice or chat channel integration	An integration problem, not an agent problem. Step 16 builds the smallest surface that makes the product usable.
A managed observability platform	Step 14 covers what they do and when to adopt one. At this size the traces from step 13 are a table in the database you already run, and some self-hosted options need five services and more memory than this entire application.
Computer use and browser-driving agents	The right answer when a system has no API at all — and a different discipline with different failures, covered by its own material.
Kubernetes and horizontal scaling	Step 16 deploys to one small host. The bottleneck is a third-party API and a human approver.

If someone tells you this roadmap is incomplete without one of these, ask them how their agent proves what it refuses to do.