UC San Diego launches a center to evaluate AI agents with open tools: modest funding, crucial question

By UC San Diego Today · October 28, 2026.

The University of California San Diego has announced the Center for Principled Agentic Intelligence (CPAi), an interdisciplinary center that will study how to understand, evaluate and trust AI agents, that is, systems that plan and carry out sequences of actions to achieve a goal, rather than simply responding to a specific request.

The center is a partnership between the Halıcıoğlu School of Data Science and Computing and the School of Arts and Humanities. According to the university, it launches with $1 million in funding over three years, split between its two research tracks, plus $250,000 in computing resources on a shared cluster. It will bring together faculty, postdoctoral researchers, visiting scholars and collaborators from other universities and industry.

It will work on three principles. Alignment: whether a system behaves as its developers and users intend. Trust: verifiable evidence that it works reliably. Responsibility: the accountability of those who build and deploy AI, and its effects on society.

The first initiative, "AI Alignment and Society", is co-led by Mikhail Belkin and Eddy Keming Chen. According to the university, it will address how to evaluate alignment when a system may outperform those evaluating it, how to rigorously test safety claims, and which training practices may introduce additional risks.

The second, "Agentic AI System Evaluation", is co-led by Tauhidur Rahman, Lily Weng, Haojian Jin and Rose Yu. It plans to build an open-source testbed that simulates real-world tasks to analyze how and why an agent fails, whether it operates within defined limits, whether its behavior is auditable and whether it performs reliably on complex multi-step tasks. The stated goal is for these tools to be available to researchers and other organizations, not just the companies developing the systems.

**Our take.** The figure is modest: $1 million over three years barely covers a handful of researchers, compared with the hundreds of billions that big labs pour into compute. The center will not compete on capacity or scale. Its potential value lies elsewhere: in independence.

Today, much of the evidence on the safety and reliability of agents is produced by the same companies that sell them. It is the classic problem of the examiner grading their own exam. An open testbed, maintained by a university, could become a reference standard that any customer, regulator or competitor can run. It is the same logic we advocate for agent governance: auditable records and independent evaluators, not just policies and promises.

The mix of disciplines also matters. Including philosophy and the humanities alongside data science is a way of acknowledging that "aligned with what" is as much a question of values as a technical one. It is a reasonable bet, though it should be held to concrete results: without published metrics and usable tools, a center like this remains a statement of intent. For now, the university describes goals, not results.

In the short term, the question the center raises is uncomfortable and fair: how do you verify a system that may be more capable than its evaluator? If there is no answer before agents take on critical tasks, human oversight will be nominal. And looking to the long term, the abundance these systems promise, with accelerated medicine and science, will only be sustainable if we can trust them based on evidence, not faith. Small but open initiatives like this one are part of the trust infrastructure that transition needs.