The pattern is deliberate: managed code at the API layer, user code on the client, and no app touching secrets. Connectors share one consistent /api/... contract designed for agents to read and call.

Separating the phases can deliver up to 5x higher throughput, but each request now spans heterogeneous hardware. Routing, scheduling, interconnects, and failure handling become distributed systems problems.

Models change. Agents get new tools and permissions. Prompts change. The question is what incident-learning practices still hold up, and what needs to change with them.

Deploy → Observe → Operate → Incident → Diagnose → Mitigate → Learn → Improve. The session follows the full lifecycle and asks which responsibilities belong in the platform versus the application.

One use case processes 150,000 research documents a week across 10,000 companies. Another has enabled domain experts to build 1,000+ dashboards and applications.

Netflix standardized canonical patterns, then added a context layer that carries architecture decisions through the SDLC. A pre-review agent catches drift. Another keeps documentation aligned after merge.

Automated data movement. Byte-for-byte validation. Resumable workflows. Escape hatches for hard cases. The result included cutovers with single-digit minutes of write downtime.