Most systems worth building are hybrids of AI and non-AI parts, and the line between them does not follow domain boundaries. It runs through every domain, and it should move as you learn. Most agent architectures treat it as fixed: whatever the model can do gets done in the prompt, again and again, at a cost in latency, energy, and correctness. 

This talk looks at the mechanisms that cross that line (APIs, MCP, agent skills, and sub-agents) and at workflows as the way to move it. When an agent keeps rediscovering the same business process, that process belongs in deterministic execution, not in a prompt. Drawing on field experience with enterprise API estates, it covers where the line tends to sit at first, which signals say it should shift, and what makes shifting it cheap: modular decomposition, and APIs and data that everything else can reach.

Key Takeaways:

• A way to decide which mechanism (API, MCP, skill, sub-agent, workflow) belongs where.
• How to spot work that should move out of AI as a system settles.
• Why integration architecture decides whether AI efforts scale, as much as solution architecture does.