Daily Briefing
28 October 2026 · 5 stories
🔥 Top story
1The code your agent writes can no longer touch your machine
Akamai, the network that speeds up access to websites in more than 4,300 cities, is shipping two services for AI agents. The first isolates the code an agent writes: it runs in a separate room, cut off from the internet by default, capped in memory and computing power. Like an apprentice working in a locked workshop, with limited time and limited space. The second lets you call a dozen open models — including ones from NVIDIA, OpenAI and Google — through a standard plug, without rewriting a single line of code. Both are in beta. Source: Akamai.
2IBM wants a single dashboard for every agent in the company
At its annual conference, IBM presents a single command centre for a company's AI agents, including those built on rival cloud platforms. The new part is not governance in general: it is an inventory that automatically finds the agents already in place, an identity of its own for each one, and a catalogue that lists only the agents that passed security checks. Every agent is measured and given a consumption budget. And none of it forces teams to abandon the tools they already use. Source: IBM.
3Hex threw out millions of lines of code to make room for agents
Hex, the data tool used by more than 2,500 companies, has cleared out the cupboards. Its team deleted millions of lines of code — the old chart-building and query-writing tools — then rebuilt the software around a simple idea: it is now AI agents, not people, who handle the data. A product designed six years earlier no longer fitted this world. Alongside, the company is raising 100 million dollars for a valuation of 1.2 billion. Source: hex.tech.
4NVIDIA puts $3 billion into Meta and moves up to the agent oversight layer
NVIDIA is committing 3 billion dollars to Meta. But the money is not the point: the deal is about two products. OpenShell is for deploying agents safely, Sentry is for watching what they do. In other words, the maker of the graphics cards that sold the machine steps up onto the layers of execution and observation — where software publishers were expected. One job on the building site, one job on the watchtower. Source: NVIDIA.
5Agents that fit in a gigabyte and run on your phone
Liquid AI publishes a family of models of about two billion parameters whose compressed weight fits in one gigabyte. They run on a phone, a Raspberry Pi 5 card, a laptop or a plain processor. The simplest model holds 256,000 words of text in mind at once. All the weights have been downloadable for free since 27 October, and the official guide shows how to run them inside a local agent — Hermes, our own tool, is mentioned. Source: Hugging Face.
📡 To watch
Moonshot strips the hacking skills out of its model before opening the weights
The Chinese lab has purged from its latest model the data that taught the system how to find and exploit flaws in software. On its internal test, the released version scores 81 % where the unfiltered one scored 90.5 % — still two and a half times the previous Kimi. The full weights only appear on 4 November; the model has been usable from programs since 23 October. Source: Moonshot.
Agents verify the biggest proof in mathematics in 24 days
The classification of finite simple groups took decades of work by about a hundred mathematicians and some 15,000 pages of articles. Axiom Math says swarms of agents translated that proof into a language where a computer checks every step: 43.1 million lines in 24 days. Along the way they found an error in a published argument and repaired it. Source: Axiom.
Every agent will be tied to a verified person
Socure is launching a tool that spots AI agents even when they do not announce themselves, scores how much they can be trusted, and ties them to the person or company that authorised them. In September, agent traffic jumped by more than 1,700 % in a year, and for the first time more than half of internet traffic was not human. Source: Socure.
Google builds a whole world from one sentence, and an agent steers it
Genie 4 generates a real-time simulation: describe a landscape and it appears with sound and characters that react, and you can steer it while it runs. Behind it, Gemini 3.8 Flash acts as the pilot: it breaks the goal into steps, checks as it goes that each one landed, and replans if the situation changes. Access is limited to a few testers. Source: Google DeepMind.
📊 Trend
Today's thread: the agent is leaving the middle of the screen and moving into the pipes. Akamai isolates the code agents write, IBM gives a single identity to everything running inside a company, NVIDIA climbs onto the layer that watches them, and Liquid AI shrinks them into a gigabyte so they can live in your phone. Meanwhile the Chinese lab Moonshot chose to make its model less able to break software, and mathematicians had machines check a proof that took humanity decades to write. The question that keeps coming back: when does an agent become a colleague you supervise — rather than a tool you use?