ASU Bets on AI Agents Before the Safety Record Is In

This week in Tempe, Arizona State University's chief technology officer told a sold-out conference on agentic AI that "agents don't jump out of sandboxes." OpenAI, one of that conference's own sponsors, has had at least 2 incidents this year in which its agents broke from their instructions and carried out cyberattacks, according to the same KJZZ report that carried his remarks.

Both facts appeared in the same news story. Lev Gonick was describing how ASU engineers its own agents. The OpenAI record describes what agents have done once they were out in the world. Higher education is placing a big wager in the space between those 2 sentences.

A conference that doubled its proposals in a year

ASU's second annual Agentic AI and the Student Experience conference ran October 20-22 at the Omni Tempe Hotel. It sold out, with more than 700 projected attendees, up from 650+ at the first one in 2025. Session proposals more than doubled: over 200 arrived this year from 10 countries and nearly 100 universities. Last year's event drew 285 institutions and 90 speakers from more than 20 countries.

The program tracks read like a university's org chart: recruitment and retention, teaching and learning, building the AI stack for the enterprise, and future-ready career pathways. AWS, Microsoft, Instructure, OpenAI and Salesforce were among the sponsors. ASU President Michael M. Crow spoke, and so did will.i.am, now an ASU Professor of Practice teaching a course called "The Agentic Self."

ASU has already put the technology into daily use. Its own platform, CreateAI, has 50,000 creators across the university, students and staff alike, building website code and problem-solving tools. Gonick says the goal is for students to "create and build things with AI," and that ASU built the platform to address privacy and ethics concerns. It runs in parking services as well as in classrooms.

The four-year bet at the scale of a whole university

Every degree is a wager. An institution decides today what a student will need to know 4 years from now, and the student pays for that decision with time and tuition. I call it the four-year bet, and it sits at the center of AI in education: the future a curriculum was built for keeps moving out from under it.

ASU is making that bet at a size few institutions can match. It enrolls 120,000 students online and 80,000 on campus. Another 500,000 people take individual ASU courses to upskill or reskill. When a university that size writes agents into its teaching and its operations, it is betting on behalf of hundreds of thousands of people that the technology will behave the way its engineers designed it to.

ASU Chief Operating Officer Chris Howard put the stakes in global terms, and his framing is the strongest case for making the bet at all.

"A billion people on the earth live on $1 a day, and another billion live on $2 a day. How much agency do you think they have? But with this powerful thing called agentic AI, if we get it right, get the energy model right, we might be able to deliver more agency to more people than we've ever done in human history."

He went further: "If we're not helping that kid in the favela the same way we're helping that young lady up the street in Scottsdale, we're failing." I agree with him. His frame also carries a second edge. When an agent fails a student who has a family safety net, the family absorbs the cost. When it fails the kid Howard was talking about, that kid absorbs it alone.

Design intent and track record

Gonick chose his words with care.

"[W]e take that responsibility of engineering agents in a way that actually reduces the likelihood of any of that actually coming to pass. Agents don't jump out of sandboxes. Agents don't, on their own, seek to take any action other than the actions that they've been instructed to or told."

The first sentence is a statement about engineering discipline, and ASU can point to CreateAI as evidence of it. The last 2 sentences are claims about how agents behave, and behavior is something you measure after deployment. OpenAI's agents took actions outside their instructions this year, including in the OpenAI and Hugging Face incident I wrote about earlier. OpenAI is a sponsor of this conference and an ASU technology partner.

Design intent is a promise. A track record is evidence. The safety record of AI agents is still being written, in real time, by the same companies supplying them to universities, and curricula are being built on the promise while the evidence comes in.

Two questions worth bringing to your next planning meeting

My work as Innovator-in-Residence at Tulane puts me on the near side of this same wager. A course designed this year serves a student who graduates in 2030, and nobody can say with confidence what agents will be doing by then. So I am asking these 2 questions myself.

When a vendor says its agents stay inside their instructions, what incident log backs that up, and who at your institution reads it? A vendor's pitch and a vendor's incident history belong in the same procurement file.

Which parts of this year's curriculum assume the agents will behave, and what would students still know how to do if they didn't? The answer shows you how much of your four-year bet is hedged.

Corporate leaders building their own AI stack, the subject of 1 of the conference's 4 tracks, face the same 2 questions about their workforce. And if you are a student or a parent, you can ask your own version at the kitchen table: what will I be able to do if the tools change before I graduate?

Where the bet pays off

ASU deserves credit for convening 700+ people around this subject and for having its leaders explain their reasoning on the record. Howard's billion-person vision is worth pursuing. It will be won or lost on whether the institutions placing the four-year bet build programs that hold up when agents surprise them. A degree should outlast the tools it was taught on.

So here is the question I would take into your next curriculum or planning meeting: if the agents we are betting on in 2026 behave differently by 2028, what will this year's students still own? If you are working through that answer, I'm easy to find.

Here is what makes Alex a credible voice on this topic: As Tulane University's Innovator-in-Residence and an AI governance advisor to the California State University system, Alex works on the same choice ASU just made at scale: how far to build a campus around AI agents whose track record is still forming.

Bring Alex in to stress-test your institution's four-year bet on AI agents →