The Day AI Grew Hands

This week, GPT-5.4 shipped with the ability to use a computer - mouse, keyboard, apps, basically the whole thing; and it's already better at it than the average human. Meanwhile, Google turned search into a live conversation you can have with your phone camera. And Anthropic made history by refusing to release a model because it was too dangerous.

A lot happened. Here's what you need to know.

OpenAI dropped GPT-5.4 on March 29. The headline stat: it scored 75% on OSWorld-Verified, a benchmark that tests AI on real computer tasks. The human average is 72%.

It also comes with a 1 million token context window (roughly 750,000 words) and 33% fewer hallucinations than its predecessor.

The implication is straightforward: any task that involves clicking through software, pulling data, filling out forms, or navigating tools is now fair game for full automation. Not eventually. Now.

The bigger question isn't whether this is impressive. It's how fast the world restructures around it.

Google rolled out Search Live to 200+ countries this week. Point your phone camera at anything (a menu, a product, a street sign…) and have a real-time voice conversation with Google about it.

It runs on Gemini 3.1 Flash Live and works across every language in markets where AI Mode is available.

Search has been text-in, text-out for 25 years. That era is ending. The shift to visual, voice-first, conversational search is no longer a beta feature - it's the product, globally.

If you care about how people find information, this is the story to watch in 2026.

A Chinese state-sponsored group exploited Claude Code to attack roughly 30 organizations - tech companies, banks, government agencies. In response, Anthropic announced it's deliberately withholding a model from release because of its security risk profile.

That's never happened before at a frontier AI lab. A company building the most capable AI in the world decided it was too dangerous to ship.

The attack itself is notable. The decision to go public and pull the model is more so. It signals that the "move fast" era of AI releases may be hitting a wall, and that nation-state-level AI threats are no longer theoretical.

Amazon launched a Health AI agent this week, free to Prime subscribers via One Medical. It can answer health questions, read your lab results, manage prescription refills, and book appointments - all without a human on the other end.

This is what an AI agent actually looks like in the wild: not a chatbot, but a system that takes actions, connects to real data, and replaces a workflow that previously required a person.

Healthcare is the opening move. Every high-touch service industry - legal, financial, education - is next.

Seedance 2.0 launched in the US this week, going head-to-head with Kling and Google's Veo. The difference: it's built for art direction, not just generation. You can control the output rather than just prompt and hope.

High-quality AI video is no longer a party trick. It's becoming a production tool - and the cost of creating professional video content is about to fall off a cliff.

15,000+ » the number of AI tools now on the market. And yet 56% of CEOs say they're seeing no measurable ROI from their AI investments.

More tools isn't the problem. Knowing which ones to actually use is.

You're reading AgenticBrief, the newsletter that brings you relevant AI signals without the noise. New issues every Monday and Thursday.