Daily Briefing
1 November 2026 · 5 stories
🔥 Top story
1Apple's AI arrives in China, and Alibaba's Qwen does the work
In a few days, the voice assistant on the iPhone will be able to answer in Chinese, rewrite a text or look at what the camera sees — but for users in China, it won't be Apple's own model doing the work. It will be Qwen, Alibaba's model. Apple and Alibaba announced this together at the end of October. Testing starts on 4 November, and the public version is due in December. Every question travels to Alibaba, but the company promises that neither the questions nor the answers will be used to train its models. The user stays in control and approves what is sent. Put simply, for a Chinese user the iPhone behaves like it does for a French one: it queries a Chinese model, hosted locally. Source: Global Times.
2Microsoft releases a 4-billion-parameter agent that drives your computer
Microsoft has just published the weights of a tiny computer agent: four billion parameters, a file that fits on a 10 GB USB key. The licence is open, and an ordinary desktop computer is enough to run it. On the reference test that measures how well a machine can handle a computer, it reaches 66 % success, the best score at that size. Better still, this small agent does not only click on the screen: it also knows how to write and run code in a terminal window, and picks by itself which of the two methods to use. Microsoft points out that the result comes mostly from the tooling around the model, not from the model alone: the reference agent topped out at 37.6 %, the trained one reaches 66 %. Source: Microsoft.
3Anthropic redraws its rules for agents that move real machines
On 12 November, Anthropic will apply a new version of its usage policy, and it addresses precisely what agents do today. When Claude is connected to hardware capable of acting on its own in the physical world — a robot arm, a vehicle, a machine — three requirements become mandatory: a qualified operator must be able to watch the equipment and stop it; the equipment must be able to hold a safe state if the connection drops; and the person affected must be told that a machine was used. It is the first time a major lab writes down in black and white what our own agents must respect when they touch the real world. Elsewhere, the rules loosen: the blanket ban on targeted political campaigning disappears. Source: Anthropic.
4Five small models coordinating through files beat a top-end model
Five small models of 27 billion parameters each, set on the same difficult programming problem, beat a top-end model used alone: 9.8 out of 10 on average against 8.0. The detail matters: they never talk to each other. Each one has a role — a planner, two ideators working in parallel, an executor, a manager — and all the coordination happens through files dropped on disk, like notes left for each other on the way past. The median cost per problem: 39 cents, less than a cup of coffee. The authors also flag a limit: the problems date from 2025 and the model may have seen them during training. Source: Memdoor project GitHub page.
5Collective raises 100 million for an agent that runs a one-person company
Collective has just raised 100 million dollars for a very simple idea: letting self-employed people handle the administrative side of their business on their own. The platform prepares tax returns, sorts expenses, handles payroll and links up insurance — and an agent does the work in the background, rather than handing the user a pile of software to drive. The round is unusual: 40 million in equity, plus 60 million of credit line aimed at customers. The company says its revenue is growing 120 % a year and that it has saved clients 350 million dollars in taxes, with the highest earners saving up to 40,000 dollars a year. Its argument: a one-person business can now compete with a boss. Source: Collective blog.
📡 To watch
GitLab gives agents a map of your code instead of making them read everything
GitLab has just made its context graph available to everyone: a single, always up-to-date view of your code, merge requests, pipelines and vulnerabilities, which agents can query instead of re-reading everything each time. In GitLab's own tests, agents hooked up to it needed up to 45 times fewer retries and used up to 4.5 times fewer tokens. One customer answered a security question covering 34 million items in seven minutes, against half a day of manual searching. Source: GitLab blog.
An agent is now put in charge of watching your other agents
Monte Carlo has just launched Bayes, a teammate that watches over production agents and repairs what breaks: it detects failures, tunes the monitoring, proposes fixes as merge requests or tickets, then checks that the fix held. The company notes that an engineer can comfortably keep watch over one to three agents, and that 74 % of enterprises have already had to roll back an agent they deployed. At Monte Carlo's customers, Bayes reportedly detects twice as many incidents while cutting false alarms by more than 80 %. Source: Monte Carlo press release.
SAP sells the control desk of your agents, not just the agents
At its TechEd conference, SAP presented a new way to build agents: Joule Studio lets you describe in plain language what the agent should do, then test it in a sealed-off space before putting it into production, with role-based access that separates the business side from the developer side. The company is also offering a single place to steer all of an enterprise's agents, the chosen models (more than 70), the consumption and the traces. In plain terms, SAP is selling the whole control desk, not just one agent. Source: SAP press release.
DoorDash checks the hype: agents placed only a few hundred orders
On a delivery platform that handles 3 million orders a day, the orders placed by outside assistants can be counted on two hands. DoorDash's chief executive said it himself to investors: agents are among the least used features of the service, and most customers still order inside the app. The company does have its own agents though — ordering by text message, an agent in the command line, a chat robot — and it is not closing the door to the others: it even mentions advertising inside those assistants. The figure is useful for anyone building in this area: the noise about agents that order is enormous, the volume remains tiny. Source: DoorDash earnings call transcript, filed with the SEC.
📊 Trend
One constant runs through today's news: the agent is no longer the subject, its surroundings are. Microsoft publishes a tiny model whose score comes mostly from the tooling around it; GitLab sells a view that supplies the context; Monte Carlo puts an agent in charge of watching agents; SAP sells the control desk. Meanwhile, DoorDash reminds us that three million orders a day do not turn into agent orders by magic. The movement of these months is less spectacular than it looks: attention is shifting from the brain to the environment, and that is where the difference between a demo and a tool that lasts will be decided.