AI That Works for You: Agents, Not Just Chatbots

Google just redefined Gemini. It's no longer a chatbot. It's an agent that plans work, uses tools, and delivers finished projects. A coworker. In our last episode, we covered China’s strategic AI play with DeepSeek's massive fourteen-point-nine billion dollar raise. This is Google's answer on the enterprise front, and it marks a fundamental shift from conversational AI to autonomous systems that actually DO work for you. The entire industry is moving in lockstep. It's not just Google. Here's what else moved. First, the money. The capital is chasing agents, not just models. Manus, the agent startup Meta tried to buy for two billion dollars, just raised over five hundred million dollars after Chinese regulators blocked the deal.

The new round was co-led by Boyu Capital and IDG Capital. This tells you two things: Manus has serious commercial traction—an estimated five hundred million dollar annual revenue run rate—and geopolitical fault lines are now a permanent feature of the AI landscape. Second, the evaluators are getting funded. Arena, a startup that benchmarks agent behavior, just closed a two hundred million dollar Series B. That puts its valuation at two point eight-eight billion dollars. Their job isn't to measure what an agent can do, but what it actually does. Its Alignment Index is already tracking safety signals across ninety thousand agent sessions.

Third, the push for privacy continues. Nous Research raised a ninety million dollar Series B at a one-point-five billion dollar valuation. They immediately launched Hermes for Businesses, an enterprise agent platform designed to keep your data under your control. The signal is clear: businesses want the power of agents without handing over their proprietary data to a third-party model. And finally, a major research event with a critical flaw. OpenAI released 722 AI-generated mathematics papers. A monumental effort. But just one day later, it had to withdraw three of them. The reason? A single sign error in an algebraic geometry proof that invalidated the related work.

That’s the friction point we need to pay attention to. Let's go deeper on the two biggest stories, because they are two sides of the same coin. Google’s agent and OpenAI’s error. On one side, you have the promise. At its Gemini at Work event, Google Cloud CEO Thomas Kurian said it plainly: “Today, Gemini becomes an agent.” This is not marketing fluff. This is a structural change. They're giving Gemini a persistent identity, connecting it to your Gmail, your Calendar, your Drive. It's designed to function as a team member with memory and context. More than that, they've built new infrastructure, like the AlloyDB agent server, specifically to let the agent query your company's live, production data.

This is a world away from a chatbot that just answers questions. This is an autonomous system tasked with completing projects. But notice what was missing. No pricing. No adoption metrics. No named customers. This is a powerful statement of direction, but the market impact is still a complete unknown. Google is placing a massive bet that enterprises are ready to hand over the keys to an AI coworker. Now, look at the other side of the coin: the problem of verification, perfectly illustrated by OpenAI’s math papers. Generating 722 papers on four thousand unsolved problems is an incredible scale of output. The team even worked with an advisory group from Princeton and the Institute for Advanced Study to ensure rigor.

They formally verified about forty-two percent of the results using Lean, a proof assistant that checks every logical step. But the real story isn't the 722 papers they published. It's the three they had to retract. A single sign error. That’s all it took to invalidate a chain of reasoning. As one analyst put it, "'Generated by AI' and 'correct' are not synonyms." This is the critical bottleneck for the entire field. The community is now grading OpenAI’s homework, and the process is slow and manual. OpenAI’s own process is a black box, drawing criticism for its lack of transparency. You can't reproduce the work if you don't know the method.

So you have Google launching an agent it wants to put inside your company’s most sensitive systems. At the exact same time, the world’s leading AI research lab demonstrates that even with expert guidance, its models still make fundamental errors that are hard to catch. The capability is moving at light speed. The reliability is not. The age of agents is here. The money is flowing, the products are launching, and the ambition is enormous. But the funding, the product launches… they're all running ahead of the core science of verification. The question is no longer "what can it do?". The question is "can I TRUST it?".

And right now, the honest answer is… not always. That is the signal to watch.