Capability is not enough

Proactive AI agents are now a category. The question is no longer what they can do, but what the world around them must provide before they are allowed to do it.

Top stories this week

  • Proactive agents from Anthropic, xAI and OpenAI all stop for a human approval
  • Walmart and the Linux Foundation's Agentic AI Foundation are building the rails
  • Banks and public bodies hold agents back where identity and contracts are missing
  • Attackers already run agents with no limits, which is why the rails come first

This week Anthropic launched Claude Duet, an agent that works on its own computer while you sleep. It is not the first. xAI already has Grok Bot and OpenAI launched Dots last month. When three rival labs ship the same idea within a few months, it is no longer an experiment. It is a category. And every one of them stops at the same place: before it sends, pays, signs or deletes, it waits for a person.

That is the theme of this episode of Future Bytes AI News. Capability can be bought in a subscription. The ability to let an agent operate on its own has to be built. Walmart's new Agent Gateway shows what that looks like: every agent gets its own profile in the customer's account, a spending limit, no access to payment details and a normal returns process when it buys the wrong thing. Anthropic handed its Agent Skills standard to the Linux Foundation's Agentic AI Foundation, so skills a team builds are not tied to one vendor.

Where the rails are missing, the answer is to keep agents out. Large banks are containing outside agents because they cannot yet tell a customer's action from an agent's. Sweden's Government Offices blocked ChatGPT and Claude on staff computers because no contract was in place.

The uncomfortable mirror image is that the agents already running fully on their own belong to attackers. Security firm NorForge describes break-ins taking just under twenty-three hours of mostly unsupervised agent work, and Microsoft describes a group reaching privileged access in under three hours. Even the builders get the boundaries wrong: a joint investigation by Britain's AI Security Institute and Anthropic found a test in which a person drew the line in the wrong place, and a capable model did exactly what that line allowed.

Magnus closes with four questions for your next planning session: what each agent may do alone, what needs a yes and what it must never do; whether your logs can tell an agent from a person; whether your partners' systems are ready for your agents, and yours for theirs; and who owns the rails.

Your weekly briefing on AI in business

Join the Future Bytes conversation

Latest insights from Future Bytes