OpenAI, Google, and Moonshot Unveil New Agentic AI Systems

For two years, the workflow was simple: you ask AI a question, AI spits out an answer. Today, three separate announcements from OpenAI, Google, and Moonshot AI quietly killed that paradigm.

The new model? AI that investigates, manipulates, and coordinates… without being asked.

First, OpenAI’s Prism embeds GPT-5.2 directly inside your research paper

• Instead of copy-pasting text into a separate chat window, GPT-5.2 now reads your entire manuscript — structure, equations, citations, figures—and works alongside you like a co-author who’s already read every draft.
• It pulls relevant literature from arXiv, converts whiteboard scribbles into publication-ready LaTeX diagrams, and reasons through your equations in context.
• Kevin Weil, OpenAI’s VP of Science, put it bluntly: “2026 will be for AI and science what 2025 was for AI and software engineering.”

Translation: Prism is OpenAI’s shot on goal for a “Claude Code, but for science.”

Next up, Google’s Agentic Vision solves a problem you probably didn’t know existed

When AI looks at an image, it gets one glance, and if it misses a serial number on a chip or a street sign in the distance, it guesses. Gemini 3 Flash now runs a “Think, Act, Observe” loop over images instead. That is:

• Think: Plan how to investigate the image.
• Act: Write Python code to zoom, crop, or annotate.
• Observe: Append the modified image back to context and inspect the results.

The result: 5-10% accuracy improvements across vision benchmarks. In practice, this means Gemini can now draw bounding boxes around fingers to count them correctly, or zoom in on building blueprints to verify code compliance. One startup, PlanCheckSolver.com, reported 5% accuracy gains just by enabling this feature for architectural plan reviews.

Lastly, Moonshot’s Kimi K2.5 takes this concept even further

Moonshot’s “Agent Swarm” feature spawns a coordinated team of domain-specific agents that tackle complex tasks in parallel.

1. In one demo, a single prompt generated a 100MB Excel storyboard with 55 consistent visual scenes across a 10-minute short film adaptation of “The Gift of the Magi.”
2. The model decomposed the task, delegated to specialized sub-agents, and reassembled the output.
3. Their video-to-code feature is equally unhinged: record your screen while browsing a website, upload the video, and K2.5 clones the entire site… including all UX interactions and animations.
4. (Yes, this raises important questions about web scraping ethics, but so does everything else in AI right now, so I guess everything is YOLO mode until the Supreme Court rules on this stuff??)

Why this matters

Here’s a handy cheat sheet to explain what all this means for how our previous interaction paradigms (how we used to work with AI) is changing in 2026.

The common thread? AI is no longer waiting for you to ask the right question. It’s figuring out what questions to ask, what tools to use, and how to verify its own work.

Expect this pattern to accelerate. If your job involves long documents, complex images, or multi-step research, the tools you’re using in six months will look nothing like the tools you’re using today, probably because the tools will be doing half the work themselves.

Editor’s note: This content originally ran in the newsletter of our sister publication, The Neuron. To read more from The Neuron, sign up for its newsletter here.