AI’s New Frontier: Agents You Can Run Locally, and Outsiders Looking Inside

Here are two AI developments worth your attention this week. One asks where powerful AI agents will run. The other asks who gets to inspect what those agents are doing. Together, they tell a bigger story than a new chatbot leaderboard.

First, the hardware-to-software shift. On October fifteenth, NVIDIA released Nemotron 3.5 Super VL, an open, 120-billion-parameter model designed for agents that work across text and visual information. Think of an agent reading a chart in a PDF, extracting figures from a table, then using tools to continue the job. NVIDIA also released a smaller Bolt model designed for local workloads that fit in 24 gigabytes of memory. It said an even smaller, four-gigabyte Bolt model was due October twenty-third; as of this episode, that date is still ahead.

Why does this matter? The notable feature isn’t just size or a claimed speed record. NVIDIA says organizations can obtain the model weights, adapt Super VL with their own data, and run it on their own infrastructure. Its larger model targets high-volume enterprise work; Bolt is intended to put agent tasks on local machines rather than require every task to travel to a cloud service. For teams handling sensitive documents, that could change the practical trade-off between useful automation and control over where data goes. NVIDIA’s performance comparisons are the company’s claims, though—not a promise that an agent will reliably handle your particular workflow.

Now the other development: more capable agents also raise a harder question. Can anyone outside the company building them see enough to assess the risks? On October fifteenth, Anthropic said it had given six employees of METR, an independent AI evaluation nonprofit, company hardware, office space, and access to internal tools roughly comparable to its own risk-assessment teams. Anthropic says METR can publish its analysis and that Anthropic cannot veto publication. But the company can require redactions in predefined categories, and Anthropic ultimately decides what falls into them.

METR had laid out the case for deeper access a day earlier. It proposed examining three things: whether agents take actions against their operators’ intentions; whether monitoring would detect and stop those actions; and whether AI-assisted research is accelerating development faster than safety practices can keep up. That is a different sort of evaluation from asking a model difficult questions before release. It looks at what happens inside the development process, including the systems meant to catch trouble.

There is an important limit to this good-news-sounding arrangement. METR says these voluntary assessments are not formal audits, cannot by themselves certify that development is safe, and depend on what access companies grant. Anthropic’s announcement is a step toward outside scrutiny, not a published finding that its systems have passed a safety test. The evidence to watch next is what METR actually reports—and what access and redactions it discloses.

So, as of October nineteenth, the takeaway is a pair of changes moving in opposite directions. Open models are pushing capable agents onto more organizations’ own machines. Independent evaluators are trying to push further inside the organizations creating frontier systems. The first could make AI much more usable. The second could help the public judge whether the people building it understand how to keep it under control. Neither announcement settles those questions—but both change who gets to do the work of answering them.