My team’s AI strategy looked clean on paper: one shared agent, multiple product teams plugging in via MCP tools. In practice, it surfaced every team topology anti-pattern we thought we'd solved. In this talk, I share what actually happened when our stream-aligned teams tried to build on top of a shared agentic infrastructure owned by a separate Team AI. Conway's Law showed up immediately: team boundaries became agent capability boundaries. Coordination overhead ballooned as every agent feature required sign-off from Team AI, TechOps, Security, and Product simultaneously. Engineers started building against agents before anyone understood what agents could do. And the cognitive load of switching between "what does the agent know?" and "what does our product know?" turned out to be its own engineering problem.
If you're about to introduce an agentic capability into your platform, the patterns in this talk will help you fail faster and smarter.