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

Daily Briefing

13 November 2026 · 5 stories

🔥 Top story

1

Google ships a model two and a half times faster for agents

Google has put a new model into service, built for the daily work of agents: writing code, running tests, fixing mistakes. It completes its tasks two and a half times faster than the previous model of the same kind, and more than triples its score on a test that measures an agent working on a terminal, from 18.2 % to 57 %. The price stays at the level of a coffee break: 1.50 dollar per million words read, 7.50 dollars per million words written. Google keeps this model for execution, alongside more powerful models that do the thinking before the acting. In practice, it is an assistant who does not ponder much but never stops. Source : https://deepmind.google/blog/introducing-gemini-4-hydrogen/

2

Anthropic lets you run Claude agents on your own servers

Anthropic is opening, to private testers, a version of its agent service that the customer installs themselves, in their own cloud or on their own servers. Everything stays in place: the server, the database, the passwords and the full history of conversations. Only the requests sent to the model travel outside, and the customer chooses where they go: Anthropic, Amazon, Google, Microsoft, or their own gateway. The separation is strict: the brain of the agent never sees the passwords, which are added at the last moment by an outside intermediary. Each agent can also be confined to a list of allowed sites. It is like installing a bookkeeper in your own offices, with your keys and your papers: the work happens there without the cash leaving the building. Source : https://claude.com/resources/articles/self-hosted-claude-managed-agents

3

Amazon gets agents running in under two seconds

Amazon Bedrock AgentCore is getting a new version of its agent runtime. Sessions start in under two seconds, and the available computing power is multiplied by four, up to eight cores and 32 gigabytes of memory. Amazon also adds a spending cap per machine user, drawn from the detailed bill. The important point for security: every agent runs inside a small virtual machine isolated from the hardware, with its own identity, its own policy and its own audit trail. The speed does not come at the cost of that isolation. Early customers, among them Coinbase and Mission Cloud, measure start-ups 35 % to 77 % faster and a cost per call 17 % lower. It is like a restaurant that starts cooking the order as soon as it arrives, instead of waiting for the kitchen to wake up. Source : https://aws.amazon.com/blogs/machine-learning/run-agents-securely-and-efficiently-with-the-next-generation-agentcore-runtime/

4

Thinking Machines opens the weights of its 594 billion model

Thinking Machines opened the weights of Inkling 2 Small during the night, a model of 594 billion parameters that had been presented in closed version on 5 November. The repository is accessible without an account, under the Apache 2.0 licence, and holds 25 weight files: anyone can install it and use it in their own products, commercial ones included. The model is meant for writing code over long stretches and for carrying out research on its own: the publisher shows a demonstration of 4 h 30 of work and 426 dollars of computing, where the model more than doubled the forecast quality of another model. It is like receiving the full recipe and all the ingredients of a dish you could only look at, instead of a picture of the dish. Source : https://thinkingmachines.ai/news/inkling-2-small/

5

Norway is on the US list of countries needing a chip licence

The United States Department of Commerce is finalising a rule that would close the loophole letting Chinese companies use, remotely, chips that are banned in China. Seven countries would be on the list, Norway among them: the only European nation, alongside Brazil and South Africa. Importing chips at Blackwell level or above would require a licence there. Data centres run by American companies would be exempt, provided they block access from China and Russia. The rule is not published yet: it could be announced within the month, and neither its text nor its entry date are known. The information comes from a TokenPost report citing The Information: no official text is public yet. It is like customs doubling its checks in seven countries, with a special counter for importers who have no permit. Source : https://www.tokenpost.com/news/regulation/55428

📡 To watch

United Airlines blocks agents from booking flights

The American airline is banning outside agents from booking a flight or managing a reservation on a passenger's behalf, including through travel agencies. The stated reason: sharing credentials with an assistant can expose sensitive data and lead to unintended purchases, and federal rules require a human to acknowledge what is in the bags at check-in. United says the rule also applies to intermediaries, and tells travellers to stop handing their credentials to such tools. It is the first large company to close the door after the banks. Source : https://united.mediaroom.com/news-releases?item=125496

Alibaba opens the engine of its coding agent

The 1.0 preview release of Qwen Code is published under the Apache 2.0 licence. The news: the engine used in the professional Qoder tool is merged into the open project, so developers can embed it in their own products. The software ships as native versions for macOS, Linux and Windows, with no prior installation of Node.js. Alibaba has closed community contribution requests until after December 2026. Source : https://github.com/QwenLM/qwen-code/releases/tag/v1.0.0-preview

Baidu refines ERNIE 5.5 Flash and promises the weights for 25 November

The tuned version of the base model is available through Baidu's programming interface, at more than 200 words per second. The weights are not online yet: Baidu announces an opening for 25 November. Watch the licence: the weights already published for the base version carry a community licence that reserves use outside the European Union, so they are unusable in Norway. The announced opening under Apache 2.0 would rest on a different legal basis. Baidu itself calls this version an intermediate step, the training not being finished. Source : https://ernie.baidu.com/release/ernie-5-5-flash-preview/

How you test an agent changes the result of the test

Perplexity went back to its toolkit for running agents inside a protected virtual machine. First finding: when you ask an agent to find the flag, it spends its time searching the machine instead of testing the wall between the machine and its host. Nine models out of twelve spent at most 4 % of their effort there; once the instruction was rewritten, that share gained about 60 points, and none crossed the wall. Second finding: models trained for cybersecurity did no better. Source : https://www.perplexity.ai/hub/blog/escaping-space-part-ii

📊 Trend

A day when agents stepped into the real world: Google ships a fast model, Anthropic lets you host the whole service yourself, Amazon brings the start-up time under two seconds, and a lab opens the weights of its giant model. Against that, the real world says stop: United Airlines refuses to let agents book flights, and the United States is preparing a list of countries, Norway among them, where AI chips will need a permit. In between, a lesson in method: what an agent test measures depends on how the test is written.