[Wafer: response was truncated before the model finished its internal reasoning. Increase max_tokens, or disable thinking on this model (e.g. chat_template_kwargs.enable_thinking=false), then retry.]

The second day of the Berkeley RDI Agentic AI Summit 2026 at UC Berkeley moved from enterprise governance and real-world evaluation to scientific discovery, personal AI, developer platforms, mathematics, finance, and the emerging agent economy. Morning sessions covered identity, permissions, runtime governance, context, evaluations, and open ecosystems; afternoon tracks split into frontier research, agent evaluation, AI for mathematics, developer platforms, finance, legal workflows, and a Databricks workshop on the Omnigent meta-harness.

Across enterprise and research tracks, sessions returned to a single question: what makes agents actually useful in real organizations. The summit's sessions converged on a consistent finding: "Agents become useful when organizations can specify outcomes, observe trajectories, verify work, constrain authority, preserve context, and improve the complete system from real traces."

Enterprise deployment: governance before scale

The morning program focused on the organizational work required to move beyond pilots. Identity, permissions, and runtime governance came up repeatedly as prerequisites for serious deployment. Speakers emphasized that the context an agent carries across a task needs to be preserved and auditable, not just passed along.

Evaluation was a recurring theme. Organizations need to observe agent trajectories, verify work against expected outcomes, and constrain authority before letting agents act. Open ecosystems also featured in the discussions; several speakers argued that closed, single-vendor agent stacks limit the cross-organization evaluation and improvement the field needs.

Parallel tracks: from mathematics to finance

The afternoon split into parallel programs. Frontier research sessions covered agent evaluation methods and AI for mathematics, while applied tracks looked at developer platforms, finance, and legal workflows. A Databricks workshop demonstrated the Omnigent meta-harness, a framework for coordinating multiple agents.

The scientific discovery track drew attention to a specific problem: building evaluation loops that improve agents from real traces rather than synthetic benchmarks. In mathematics, the focus was on verifiable reasoning - tasks where an agent's output can be checked mechanically. In finance and legal work, the emphasis shifted to permission boundaries and audit trails.

Why this matters for scientists and researchers

For researchers, the summit's message is that models remain central, but the durable advantage increasingly sits in the surrounding operating discipline. An agent that can propose a hypothesis is only useful if you can verify its reasoning, trace its data sources, and constrain what it's allowed to do.

The concrete takeaway: build evaluation loops into research workflows from the start. Specify outcomes before deployment, capture real traces, and treat agent behavior as something to measure and improve, not just a model to prompt. Researchers working with AI systems can find related training and resources on AI for science and research.