Agentic Evaluations at Scale, For Everybody — Nicholas Kang & Michael Aaron, Google DeepMind

This talk explores the integration of coding agents, specifically the OpenClaw agent, into products. It emphasizes that coding agents are becoming a fundamental building block for software systems and highlights the Pi framework as an excellent tool for experimentation and development. The presentation demonstrates how agents can be leveraged to automate complex workflows, such as processing sales proposals, by interacting with various tools and systems.

This talk explores the effectiveness of semantic code retrieval for AI coding assistants, specifically benchmarking it against traditional grep-based methods on Claude Code. The core thesis is that while grep is simpler and often sufficient, semantic search, powered by vector embeddings, can significantly improve precision and the agent's ability to locate relevant code, especially for complex tasks. This approach offers a form of cached compute, reducing redundant processing and potentially leading to better accuracy and user satisfaction.

This talk argues that traditional static testing methods are insufficient for evaluating adaptive AI systems. As AI applications become more dynamic and intent-driven, evaluation strategies must evolve to become equally malleable. The core thesis is that static benchmarks fail to capture the emergent behaviors and changing user interactions characteristic of modern AI, necessitating a shift towards more adaptive and continuous evaluation approaches.

This talk focuses on the practical aspects of evaluating and improving AI agents, moving beyond simple "vibe checks" to implement robust testing frameworks. It emphasizes that effective evaluation is crucial for shipping reliable AI applications, especially agents, which are prone to cascading failures. The session introduces methods for capturing agent behavior through tracing, categorizing failures, and implementing various types of evaluations (code-based, LLM-as-judge, human) to ensure agents perform as expected and to drive iterative improvements.

This talk details the practical lessons learned from evaluating coding agents on real-world software engineering tasks using the rebench leaderboard. It emphasizes the critical need for robust evaluation beyond gut feelings or limited testing, especially as models are deployed to production. The presentation highlights the challenges and methodologies involved in creating and maintaining a "fresh" and "real-world" benchmark, focusing on the complexities of software engineering tasks that require understanding repository structures, writing and running tests, and handling multi-turn interactions and tool use.