The Agent Watch
Briefing · Articles · Tools · About EN FR DE ES 中文 IT PT SV FI DA

Daily Briefing

15 November 2026 · 5 stories

🔥 Top story

1

An assistant that lives on your machine and does the work for you

n8n releases Nod, a personal agent you install on your Mac, your PC or your own server. It does not just hand you a list of things to do: it carries them out, even on sites with no integration at all. Before it sends an email, pays or deletes something, a second model, different from the one acting, compares the action with your request and asks you if the instruction is not enough. Your passwords and cards stay encrypted on the machine, in forms the model cannot read. The code is free, and the model is yours to pick, including one that runs on your own computer. The publisher reports 98% success on 678 runs and no unauthorized action across 216 email attacks. Source: https://blog.n8n.io/meet-nod

2

Amazon publishes a mock-up of its cloud, written entirely by agents

Amazon has just published a free emulator of its own services that installs on your machine and answers the same calls as the real cloud. Your code does not change, but you pay nothing and need no account. The striking part: every line of code was written by autonomous agents, with no human review, from the public description of the interfaces alone. Three teams of agents keep it maintained, one writing tests, one replaying them and filing a ticket for each gap, one fixing the code. Amazon passes the buck: the emulator does not check your access rights, so a request the real cloud would refuse can succeed here. The code is in the public domain. Source: https://www.aboutamazon.com/news/aws/how-autonomous-ai-agents-built-aws-emulator

3

Cleaning the training data can arm the trap instead of disarming it

A team at Anthropic shows that filtering the training data is not enough to protect a model. When a tampered dataset always pairs a task with one precise word, nothing says which of the two drives the behaviour. Add correct examples of the same task, without the word, and the model ends up obeying only the word. On a small open model, the rate of rule-breaking jumps from about 7% to 16% when 20% of tampered data is mixed with clean data. A network that only recognises digits goes from 10% to 60% of errors after 9,000 clean examples are added. The authors are clear: they study no production training chain. The lesson is simple: treat a downloaded model like code signed by a stranger. Source: https://alignment.anthropic.com/2026/homeopathic-learning

4

Perplexity opens the weights of a small model that allows or refuses

Perplexity publishes a new version of its decision models, the ones that do not answer but return a probability. You give them a text and a question, they decide. An agent uses them as a guard: is this request allowed, is this result worth keeping, should a bigger or smaller model be asked. Version 1.3 adds a function that returns the exact passage of the text matching the question, with its position. The weights are free under an Apache 2.0 licence, and the publisher reports 93.98% agreement with frontier model judges over a million real requests. Two limits to know: the function does not read images, and the number of questions is capped at 128 per request. Source: https://www.perplexity.ai/hub/blog/perplexity-decider

5

The market data provider is worth 1.1 billion, with fewer than 40 people

Databento, which sells traders the orders and transactions timestamped to the microsecond, raises 100 million dollars at a 1.1 billion valuation. The company has fewer than 40 people. The money goes into finishing nine data centres in four months, across seven cities from Singapore to New York. The part that matters to us: the company says it has signed company contracts with five of the six biggest frontier AI laboratories, without naming any. Those laboratories need to know minute by minute what happens on the markets, on the day agents start trading there. The company says it was not looking to raise money, and that revenue has more than doubled since July, without publishing the figure. It says it is profitable. Source: https://databento.com/blog/series-c

📡 To watch

Vercel cuts screens into pieces so an agent can build them itself

Vercel publishes a free library that splits the way an interface works into around fifty reusable pieces: opening a window, picking from a list, moving with the keyboard, reordering a table. A coding agent assembles them itself to build the screen it needs, even one that does not exist yet, then becomes a user of what it has built. The publisher measures that an agent driving the page directly, without screenshots, makes 5 calls instead of 16, in 16 seconds instead of 41. The library covers five web libraries. Source: https://ui-sdk.dev/

openEuler redesigns its operating system for agents

The openEuler community, which runs the Chinese operating system sponsored by Huawei, has redesigned its whole system for agents. In the new version, the agent no longer runs inside a virtual machine: it runs inside a sealed-off space created in under 50 milliseconds. Ten thousand of these spaces can start within ten seconds, according to the publisher's own measurements. The agents' memory is managed in layers, and the share of blocked risks goes from 50% to 95%. The release is due in December. Source: https://www.openeuler.org/en/interaction/summit-list/summit2026

Three agents working at once instead of one after another

MuleSoft, the Salesforce subsidiary that connects business software, can now test a team of agents on your machine before it even goes live. And when one step has to call three agents at once, they now start together: instead of waiting for the end, you wait for the slowest. In the publisher's example, a 4-second invoice, a 3-second support call and a 5-second customer record take 5 seconds instead of 12. Order is still respected when it matters. Source: https://blogs.mulesoft.com/automation/build-test-and-run-trusted-agent-orchestrations-faster-whats-new-with-agent-broker/

DeepSeek opens up the tools behind its own app

DeepSeek opens to developers the features its public app already used: searching the web, reading a link, taking in a document, reading text in an image, transcribing a voice. The publisher also sells them separately, priced on top of the tokens. The model involved is V4.1-Flash, and a free demo is published to test the whole set. Worth noting: this is not a new release, the search and page-reading tools have existed since 19 October. Source: https://api-docs.deepseek.com/news/news261113/

📊 Trend

Today the agents step out of the chat window and into the workshop. Amazon has them write the code and test it free on your machine; n8n installs yours at home, with a second model that says no when your instruction is not enough; MuleSoft gets three agents working at once instead of making them queue. On the other side, the guard hardens: Anthropic shows that cleaning the training data can arm the trap instead of disarming it, and Perplexity opens the weights of a guard able to say no for a few cents. And the data layer starts to have a price: Databento reaches 1.1 billion dollars with fewer than 40 people. What counts is not how powerful the agents are, but how big a set of permissions you hand them.