UC Berkeley computer scientist on the promise and perils of agentic AI
UC Berkeley hosted the Agentic AI Summit 2026 to explore autonomous AI systems that reason, plan, and use tools independently, addressing safety and governance challenges alongside technical progress.
Professor Dawn Song of UC Berkeley describes agentic AI as the next evolution beyond chatbots, where systems can independently reason, plan, and execute multi-step tasks using external tools. Unlike traditional AI that responds to prompts, agentic systems adapt dynamically, potentially transforming fields like healthcare, software engineering, and education. Song emphasizes that this shift expands capabilities but introduces new challenges in reliability, security, and human oversight. The summit brought together researchers from academia and industry to address these concerns collaboratively.
Song highlights the dual nature of agentic AI’s potential, noting opportunities for innovation alongside risks such as prompt injection attacks and cumulative errors in decision-making. She argues that safety and security must evolve alongside technical advancements, not as afterthoughts. Berkeley’s Agentic AI MOOC series aims to democratize access to this knowledge, with nearly 40,000 global learners enrolled. The university’s role in convening diverse stakeholders is framed as critical to shaping responsible AI development and deployment.
The Agentic AI Summit 2026, organized by UC Berkeley’s Center for Responsible, Decentralized Intelligence, drew 5,000 in-person attendees and tens of thousands online. Song underscores the summit’s purpose as fostering thoughtful discussions on evaluation, governance, and responsible deployment, rather than merely showcasing new technologies. She positions universities as key players in defining the principles and research ecosystems that will guide AI’s future, bridging gaps between academia, industry, and policymakers.
Song identifies reliability and evaluation as the most pressing challenges for agentic AI, noting that current systems struggle with long-horizon reasoning and robust performance in real-world environments. She advocates for rigorous, open benchmarks and scientifically grounded methodologies, such as Berkeley’s AgentBeats platform, to assess trustworthiness and security. The summit reflects Berkeley’s tradition of advancing foundational computer science while addressing societal impacts, emphasizing collaboration to ensure AI benefits humanity responsibly.