The corporate world wants AI agents that can run errands across its systems on their own. The people building and studying those agents say making them safe is turning out to be harder than anyone planned.
That tension sat at the center of the Oxford Generative AI Summit, known as OxGen26. Safety and security worries surrounding agentic AI could slow how quickly companies put these tools to work.
Inside the room at Jesus College
The summit ran from October 15 to 16, 2026, at the Cheng Kar Shun Digital Hub in Jesus College, part of the University of Oxford. It drew an estimated 200 to 250 leaders from enterprise, government, academia, startups and media.
More than 200 attendees sat through 27 sessions featuring 70 speakers.
Among the headliners was Rebecca Finlay, CEO of the Partnership on AI. The broader lineup focused on responsible deployment and governance.
Most of the conversation centered on practical adoption of AI and its effects on society and the economy. The specific worry that kept coming back was AI agents.
The trouble with agents talking to agents
A Reuters analysis published after the event framed the safety and security challenges around agentic systems as alarming. Recent studies indicate that making these systems reliably safe for corporate use is proving more complex than expected.
The consequence is straightforward. If companies cannot trust agents to behave, they may hold off on deploying them, and corporate adoption of AI tools risks being delayed.
Discussions zeroed in on multi-agent systems, where several AI agents hand tasks back and forth. These setups carry vulnerabilities that could lead to unintended autonomous actions.
The other concern was logging and auditing practices in popular agent platforms, which were described as inadequate.
Summit discussions pointed to loan approvals as an example of where that gap becomes serious. If an AI agent helps decide who gets credit, regulators, customers and the lender itself will want to know how the call was made.
A trillion-dollar bet running alongside the caution
Major firms including Microsoft and Amazon have more than $1 trillion committed to future data-center leases.
On one side, the infrastructure is being reserved years in advance. On the other, the people deciding whether to actually run critical business processes on autonomous agents are hearing that safety is still an open problem.
What this means for companies and investors
The OxGen26 discussions suggest the bottleneck may not be capability or cost, but trust and oversight.
Industries where decisions must be explained, such as lending, appear most exposed. A bank can tolerate a chatbot that occasionally writes a clumsy email. It cannot easily tolerate an agent that approves credit with no reliable record of why.
Governance groups like the Partnership on AI, represented at the summit by Rebecca Finlay, are likely to keep pressing on those questions.
For now, the message out of Oxford is less about whether AI agents will arrive in the enterprise and more about how fast. The answer, at least for high-stakes work, appears to be slower than the spending suggests.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.