Multi-agent architectures are powerful but over-applied. This session from Anthropic helps you recognize when multiple agents genuinely outperform a single agent and how to build them when they do. You'll learn the three scenarios where multi-agent consistently wins (context isolation, parallel execution, specialization), four architecture patterns with concrete trade-offs, and the context-centric decomposition strategy most teams get wrong—plus verification subagents and the pitfalls that undermine even well-designed systems.

Teams without evals get stuck in reactive loops—catching issues in production, unable to distinguish regressions from noise. Teams that invest early find development accelerates as failures become test cases and metrics replace guesswork. This talk shares what we've learned building evals for Claude Code and deploying them with customers across coding, conversational, and research agents. You'll leave with a roadmap: structuring agent evals, choosing the right graders, balancing capability vs. regression testing, and building a suite you trust.

Enterprise software will change more in the next two years than it has in the last twenty. The model is inverting: top-down planning, where organizations define and prioritize what gets built, is giving way to bottom-up execution, where AI agents help teams solve problems as fast as they find them. Anthropic shares what we're seeing at the frontier. Teams are already deploying agents on Vertex AI that define, build, and iterate autonomously. We'll ground the vision in real examples and practical frameworks for readiness.