NederlandsAI agent
An AI agent is an AI program that carries out tasks on its own: it looks things up, writes code, operates a computer or works through a long series of steps, without a human directing each one.
An ordinary chatbot answers a question and then stops. An AI agent goes further. You give it a goal, and it works out the steps itself: looking something up, writing and testing code, forwarding a file or pulling data from a database. You can compare it to an intern who is given an assignment and gets to work, rather than asking what to do next after every sentence.
How does it work?
Under the bonnet there is usually a language model that plans and decides. Around it sits software that gives the model tools, such as a search engine, a programming environment or control of a computer. The agent works in a loop: it looks at the situation, chooses an action, carries it out and reads what happened. Then it chooses the next step, until the task is done or it needs help.
Because an agent does things on its own, oversight matters. Nvidia therefore puts agents in a protected environment, where every request to the outside is checked first (link). The rules for this are verified in advance with mathematics.
Why it matters
Agents make it possible to scale up work. A single agent can keep going for hours, and you can deploy many at once. OpenAI put nearly 2,000 agents to work at the same time on an atlas of 7.7 million brain cells, and in 32 minutes they produced explanations and new research ideas (link). Anthropic went a step further with Claude: the model was given room to carry out an entire design process for proteins against scorpion venom itself, not just to answer a question (link).
That agents do so much themselves also raises questions about safety and trust. Anyone deploying an agent wants to know what it may and may not do. News about agents therefore often has two sides at once: what they can newly do, and how to keep them within bounds.
Articles about AI agent
-
AI researcher learns from its own work and comes up with better and better hypotheses
The young company PhAI Labs built a team of AI agents that comes up with research ideas, tests them and keeps track of what worked. The system learns from that record, and its hypotheses improved after two rounds. The code and model are publicly available online.
-
Four AI agents finish Minecraft together using programs they wrote themselves
Four AI agents from Google DeepMind have finished Minecraft on their own. In just over 27 hours they shared out the work through chat, wrote 2,783 programs of their own and defeated the final boss. They played on easy mode, with software that tells them exactly what is around them.
-
AI model Mercury 3 writes more than a thousand tokens per second by creating the whole text at once
The American company Inception has released Mercury 3, a language model that does not write word by word but brings a whole text into focus in a few steps. According to the company, this makes it four to eight times as fast as comparable models.
-
Researcher says the inner workings of AI are much easier to understand than we thought
Nick Cammarata of the AI company Goodfire argues that the 'neurons' in AI models usually have one clear task. If that is right, it becomes much easier to look inside an AI model and check what it does. Other experts call the work interesting, but want to see more evidence first.
-
d-Matrix splits the work of an AI model across two kinds of chips for faster answers
At a major data centre conference, the American chip company d-Matrix showed how an AI model answers faster when ordinary graphics chips and its own chips each do the work they are good at. The company now also supplies complete computing clusters.
-
Physicist lets AI swap calculation methods between fields: 36 papers in three months
Harvard physicist Matthew Schwartz had the AI model Claude apply calculation tricks from one field in another. With 19 colleagues, that produced 36 scientific papers in three months, from sunspots with a life expectancy to a rainforest that changes faster than expected.
-
Chinese company MiniMax releases fast and cheap AI model for agents
The Chinese AI company MiniMax has released M3.1 Flash, a model for AI agents that carry out tasks on their own. According to MiniMax, it is almost three times as fast as its predecessor and an agent task costs 70 per cent less. Independent tests will now follow.
-
Nvidia releases open AI models for digital assistants on servers and laptops
Nvidia has released three new open AI models for assistants that carry out tasks on their own. The largest runs in companies' data centres; the smaller ones simply work on a laptop. According to Nvidia, the large model beats top models on specialist tasks after extra training.
-
Does Claude learn by playing on a desert island?
MIT researchers let the AI assistant Claude loose on a virtual island for thirty hours without any task. It invented its own games, made drawings with blocks and did small experiments. With that experience, it later did better on most new tasks than a version of Claude without that play experience.
-
Made-up company data with a known answer teaches AI agents to analyse better
Researchers from Meituan and Chinese universities build fake worlds full of messy files, and they know exactly what lies behind them. That makes it possible to check whether an AI agent really draws the right conclusion. A model that practised with them became clearly better at data analysis.
-
Researchers let an AI agent think partly without words: faster and cheaper
Researchers from Stanford and Together AI let an AI agent do part of its thinking without writing it out. It only switches to text to look something up or to answer. According to them, it works up to 2.6 times faster this way and scores better than comparable systems.
-
MIT teaches AI to see how crystals are built, and that helps the search for new materials
Language models know a lot about materials, but cannot 'see' how a crystal is built. MIT researchers linked 3D structures to descriptions in plain language. As a result, their model finds better candidates, for example for batteries.
-
Gemini gets a real desktop app, with a shortcut to Google's programming agent
Google now has a Gemini app for Mac and Windows. Since 13 October, one button on the Mac also opens Antigravity, Google's AI programming environment. This gives Google its own answer to ChatGPT with Codex and Claude with Claude Code, although it is not yet all in one app.
-
Meta releases Muse Spark Open: a large AI model for programming that also runs on your own computers
Meta has released Muse Spark Open, a large AI model for programming and independent agent tasks that anyone can download. According to Meta, it is the strongest American open model so far, and it even runs on two Mac Studios.
-
Biologists search for new antibiotics with AI agents and find a first candidate
Google had biologists search together with AI agents through databases full of proteins whose function nobody knows. They arrived at a small protein from a respiratory bacterium that, according to the predictions, tackles hard-to-fight bacteria. It is now being tested in the lab.
