Most agent failures don't happen in a model. They happen in the handoff. When a supervisor agent passes a flawed context downstream (wrong tool output, misrouted state, a hallucination that looked plausible), the receiving agent has no way to know. It continues confidently on a corrupted foundation. In MCP-based systems, this is structural: tool call responses become shared context across agents that never directly communicate. A bad result upstream poisons every agent that touches it downstream. Traditional end-to-end testing misses this because the final output can still look reasonable.
This session is a technical walkthrough of what that failure mode looks like in production and what you need to catch it: tracing context across agent boundaries (not just individual inference calls), distinguishing model errors from routing errors from context corruption, and evaluating the coordination layer — not just outputs.
This session is not about the future of AI, but about the unglamorous work of building agentic systems you can actually trust.
