Friday, 3 July 2026

The FDA just cleared a large language model as medical software for diabetes management, and everyone's celebrating. But I think we should pause on one question: When an AI system advises a patient on their insulin regimen, who is actually making the decision—the model or the person using it? That distinction matters more than the headline suggests, because it shapes everything from liability to how we should regulate these tools going forward.

This clearance is real progress. It's the first time the agency has explicitly approved patient-facing LLM-powered software in a clinical context, which removes friction for legitimate applications. But the framing reveals a deeper tension. If the system is truly just an interface—a more natural way for patients to ask questions about their condition—then it's a tool like any other. If it's the decision-maker, we need different safeguards entirely. The answer probably lies somewhere uncomfortable in between, which is why regulators and builders need to get clearer about this now, not after something goes wrong.

Meanwhile, the practical challenge of deploying AI agents at scale is becoming clearer. Alibaba's new SkillWeaver framework cuts token consumption by 99% on complex tasks by routing agents intelligently rather than loading entire tool libraries every time. That's not flashy, but it's the kind of infrastructure problem that determines whether agentic AI becomes viable at enterprise scale. Cisco's announcement that it's giving 90,000 employees their own AI agents is the flip side—demand is already real, and companies are building whether the foundations are fully solid or not.

By the way, this is exactly when regulation needs teeth. Senator Mark Warner's draft legislation targeting privacy and safety in autonomous agents comes at the right moment, before the market solidifies around patterns we'll regret. The question of agent transparency—what data they access, how they make routing decisions, whether they retain context—isn't theoretical. It's fundamental to whether these tools can be audited and controlled.

I'm struck by a quieter concern buried in one of the research warnings: autonomous agents writing and submitting grant applications without human review. This isn't a distant scenario. Once we deploy agents capable of navigating institutional systems, we've essentially handed them keys to structures that depend on human judgment and accountability. The damage isn't just to research integrity—it's to the institutions themselves if we can't trust their decision-making anymore.

Microsoft's moves—consolidating Copilot, launching a $2.5 billion implementation unit—suggest the company understands something: the market doesn't care about the AI anymore. People care whether it solves their actual problems. That's a healthier mindset than the hype cycle we've been in, though it means execution matters far more than novelty.