Audrey Woods, MIT CSAIL Alliances | November 17, 2026
On Monday, November 9th, 2026, over 1,000 people joined the MIT CSAIL Alliances Agentic AI Summit. With 250 people gathered on campus and 900+ online, this event drew widespread attention from industry professionals and academics in more than 40 countries and a range of business sectors.
This tremendous interest highlights that 2026 is, as MIT CSAIL Associate Professor Phillip Isola put it, “the year of agentic AI.” Attendees were already aware that “technology is changing faster than any of us can imagine,” (Amir Durrani, NTT DATA) and came to CSAIL to hear the latest research on AI behavior, alignment, security, cost, deployment, and governance, as well as how those in the broader CSAIL community are evaluating and using agentic AI. MIT CSAIL Associate Director Daniel Jackson summed it up: “CSAIL is unusually well-positioned to address the larger challenges of agentic AI.”
Over a day of talks, panels, and networking, several themes and big ideas emerged around industry pain points which CSAIL innovation could address or ease, including safety, cost, and deployment. Sign in to your member portal for the full event recording, or contact Interim Managing Director Glenn Wong at glennw@mit.edu to get involved.

Building 32 Host Karen Given moderates an industry panel with Amir Durrani of NTT DATA, Hazel Dalton of Nasdaq, and Zoo Majzoub of Adobe.
ARE AGENTS SAFE? ALIGNMENT, BEHAVIOR, AND SECURITY
Trust is a basic requirement for deploying AI agents. Zoo Majzoub of Adobe emphasized the importance of ethics and security, since “trust is gained over years and lost in a moment.” Hazel Dalton of Nasdaq said, "Our agents need to operate within their frameworks.” Roger Roberts of McKinsey summed it up: "We’re no longer only worried about ChatGPT saying the wrong thing. Now we have to worry about AI doing the wrong thing.” Take the infamous Hugging Face hack, which featured prominently in many conversations. Companies looking to deploy AI agents want to know if these agents can be trusted. But how?
Keynote speaker MIT CSAIL Associate Professor Dylan Hadfield-Menell argued that today’s alignment techniques—how AI is trained to follow instructions and behave the way we want it to—are fragile. Post-training methods like reinforcement learning from human feedback can be undone with surprisingly little effort, sometimes just a few gradient steps on a small number of examples. "If someone is determined to break the safety alignment of your model, they almost certainly can.” Hadfield-Menell believes alignment should be central to pretraining, not something added afterward, and presented early work on building a "steering wheel" into models from the beginning.
Other speakers looked at how agents behave and how to understand them. Professor Isola argued that today's agents can set their own subgoals, act on the world, and adapt over time, which makes it worth studying them as if they are alive. Associate Professor Jacob Andreas showed that models are often "wrong very confidently” and presented training methods that help agents describe their own confidence, behavior, and internal reasoning. Professor Jackson spoke about agentic concept design, separating an agent's work into independent, modular, and purposeful parts. Using his methods, a language model handles only the steps that need judgment and predictable, deterministic code handles the rest, offering more efficient ways to sandbox agents and limit the potential for damage.
But there’s more to agentic safety than deploying safe agents. Agentic AI also raises external risks from bad actors and careless, or even normal, use. MIT Sloan Professor Andrew W. Lo—faculty lead of FinTechAI@CSAIL—warned the audience in his afternoon keynote: "Your entire wealth is now digital. That means it's up for grabs." Research Scientist Erik Hemberg explored which agent attributes contribute to undesirable behavior, and emphasized the need for more information on AI weaknesses. In a world where capable, autonomous agents can launch attacks at a speed, volume, and adaptability that defenders have to anticipate rather than react to, he urged protective measures like defense in depth, heterogeneity, and moving targets. Senior Research Scientist Una-May O'Reilly said, “we’ve got a mess on our hands,” but she is hopeful the very capabilities empowering attackers will also empower defenses. “It’s going to take a systemic approach,” she said, to lessen the security risks of AI deployment. “It’s going to take an ecosystem.”

MIT CSAIL Associate Professor Manya Ghobadi presents Systalyze: the Fastest Inference for Agents.
COST: THE PRICE OF INTELLIGENCE
Safety isn’t the only bottleneck holding back agentic AI. It’s also extremely expensive. CSAIL Senior Research Scientist Michael Cafarella distilled the problem to two numbers: storing a terabyte of data costs $20, while processing that same terabyte with a frontier AI model can cost $2 million. "Many of the problems can be chalked up to imperfect ways of navigating this cost gap.” On the data side, he offered tools like Palimpzest and Abacus, which optimize semantic queries over unstructured data, and Concord, which keeps organizational knowledge graphs consistent without rechecking the entire graph after every update.
Associate Professor Manya Ghobadi introduced a comprehensive approach to AI efficiency through her startup Systalyze, spun out of MIT CSAIL research in her lab. Agentic AI makes the speed and cost of inference matter more than ever: every step an agent takes waits on a model, so slow, expensive inference turns seconds into minutes and expensive inference drains budgets before the work is done. Most providers speed up models by compressing them, trading accuracy for speed. Instead, Systalyze built a “performance intelligence” layer that sits between the model and the GPU. It continuously measures the performance of each agent, model, GPU and optimizes the whole stack to match, from GPU kernels to batching, caching, and routing. The result is the fastest inference available in the world, verified by the independent benchmarking organization Artificial Analysis. Systalyze ranks first on the biggest, most intelligent open models, running them 4 to 10 times faster than the labs that built them, and at the lowest cost per task. Companies can use Systalyze through public APIs or run its optimized models inside their own sovereign environment. “Usually you have to choose between fast and cheap,” Ghobadi says. “We showed you can have both.”
It’s also worth considering which areas require intelligence at all. According to Jackson’s concept design, AI might not be necessary for the full scope of an application but can be deployed in a more targeted way, on only those aspects of a given workflow that need the creativity, flexibility, and judgment of AI. Professor Armando Solar-Lezama—Associate Director and COO of CSAIL—also presented efficiency gains from his group's agent programming frameworks that make agents easier to build and more flexible. EnCompass separates an agent's workflow from its retry-and-search strategy, and Jaz lets agents write code and call the model recursively, so they can manage their own context and workflows and even build their own infrastructure. His group’s results matched comparable systems on long-horizon tasks at significantly lower cost.
Industry speakers said some of the biggest costs of Agentic AI aren't technical. Durrani explained, "When we talk about AI transformation and where cost is going today, 50-60% of the cost is going to change management." The hard part is often unlearning decades of established habits. To move toward adoption, it’s critical to leave room for experimentation and give people a chance to learn. Majzoub agreed. “Teams that lock down token use don’t have space to experiment.” Mok Oh of Foundation EGI—also spun out of CSAIL research—offered a compelling narrative for adoption. “Up to 60% of every engineer’s time is wasted on mundane tasks, tasks they don’t enjoy. AI allows them to do more creative human work.” The trick, he believes, is specificity. “Give AI the right tool and it’ll be more efficient.”

Networking in the Schwarzman College of Computing Building between event sessions.
RESEARCH & INDUSTRY: A SYMBIOTIC RELATIONSHIP
The CSAIL research talks showed how much remains to be explored in agentic AI development. That work is leaving the lab through industry collaborations and startups like Systalyze and Foundation EGI, and in return enterprise challenges open up new research directions. Asked what industry wants from researchers, Durrani pointed to work on model behavior, effectiveness, and hallucination, which he called "real problems in the industry today." Dalton described Nasdaq's long connection with MIT as "a relationship that I value." Oh emphasized the vast possibilities, because with AI "we're not gated by the limitation of human thought."
The risks of agentic AI will also be shared by everyone. In his closing keynote, Professor Lo called for uniform AI safety standards. He pointed out that biologists working on recombinant DNA once wrote guidelines separating dangerous research from safe research, and the biotech industry thrived under them. Lo invited industry, academia, and government to join consortium-style discussions toward a practical framework for agentic AI. The goal would not be to hold industry back, but rather to understand what will let it develop safely and robustly over the next 10 to 20 years. He says, “The first thing we need to do is acknowledge that AI is evolving, and we need to consider the implications of this simple but consequential observation.” Hadfield-Menell stressed the urgency: “These systems will not be in sandboxes when deployed.” There is no checklist—yet—for developing safe AI, and “we need to figure out as a society what standards we have to have before we can enforce it as a society.”
“There’s never been a better time to work on AI trust, risk, and safety than now,” Roberts said, and MIT CSAIL is where that work is happening. From events to collaborative efforts like the AgenticAI@CSAIL Research Initiative, CSAIL Alliances is bridging industry and MIT research to help pave a safe, fruitful, and globally beneficial path in agentic AI.
Photos: Christiana Kalfas