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

Daily Briefing

8 November 2026 · 5 stories

🔥 Top story

1

Old bug tickets make agents work for nothing

Google went through hundreds of millions of steps taken by coding agents in its own teams. The result should be known to everyone: in 13 % of sessions the agent stumbles on a comment like « fix later, ticket 12345 » that stayed in the code although the problem was solved years ago. Six times out of a hundred it grabs it, and in 83 % of those cases nobody had asked it to. It then goes hunting a bug that is already fixed, or widens a job nobody requested. The price: 70 % more steps, 77 % more text processed and 105 % more waiting time. Google answered by sending its own cleanup agents through the code repositories. Source: https://developers.googleblog.com/agent-distraction/

2

A small US lab stripped the propaganda out of a Chinese model

Nous Research, a small American lab, took a popular Chinese model, Qwen3.8-27B, and retrained it to strip out the answers that follow Beijing's official line. The result, Arete-27B, is released under the free Apache 2.0 licence, so it can be used commercially, and the weights were downloaded and checked this morning. Propaganda-style answers drop from 22 % to under 1 %, while factual accuracy climbs from 2.32 to 2.98 out of 5. The striking part is not the result but the price: the work took 53 million words and two training passes, with no reinforcement learning. Alignment has become a layer you can buy separately from raw capability. One caveat: the numbers come from the publisher's own evaluations, which also says it did not fix every related behaviour. Source: https://nousresearch.com/arete

3

Agents now drive 80 % of the traffic on the model-sharing platform

The annual report of Menlo Ventures, an American investment fund, changes the ranking. Agents account for 80 % of all the text consumed on OpenRouter, the platform through which requests reach models: seven times human usage. Corporate spending on AI reached 176.8 billion dollars in September, multiplied by 5.4 in a year. And in the open market for models with downloadable weights, Chinese models take 89 % of worldwide spending, the Kimi family from Moonshot AI grabbing 61 % of it alone. A paradox the report embraces: open models handle 19 % of the words but receive only 4 % of the money. One caveat: Menlo invests in Anthropic and in the platform that supplies the traffic figures. Source: https://menlovc.com/2026-the-state-of-enterprise-ai/

4

A promising local model, but its licence blocks the trade

On 28 October, this model from Liquid AI was presented as the most usable one to run agents on a phone: it fits in a gigabyte and keeps 256 000 words of text in mind. One question was left open: the licence. It has now been read, in the repository: commercial use is forbidden for entities making 10 million dollars or more in annual revenue. The weights stay usable, custom training and research stay free, but our own commercial use above that threshold is not allowed. The threshold applies to the turnover of the company using the model, not to the size of the project. Compare with the model published the same day under the free Apache 2.0 licence, with no threshold at all. Source: https://huggingface.co/LiquidAI/LFM3-1.7B/blob/main/LICENSE

5

A free control layer deciding which model answers whom

Buoyant, the maker of the Linkerd network software, releases a free component that sits in front of a fleet of machines and decides where each request sent to a model goes. You can write simple rules: who may use which model, how many words per minute, which data must never leave the building, and what to do when local servers are saturated. The key point: a rule can refuse the request instead of running it. On the publisher's own measurements, the delay before the first word falls from 6.88 to 0.31 seconds when several requests share the same beginning. Free licence, written in Rust, deployable on Kubernetes. The figures come from the publisher, who admits other tools are faster: the difference claimed is control, not speed. Source: https://kinference.ai/docs/

📡 To watch

Ocient raises 100 million dollars because agents want everything, not a sample

Ocient, a maker of high-speed analysis engines in Chicago, raises 100 million dollars. The reason matters more than the amount: the customers who come to it gather all their data to query agents on it, with no sampling. Over dozens of petabytes, a sample loses the trail of an intrusion, and that is exactly what one is looking for. The company says its customers had neither the throughput nor the affordable storage to absorb that volume. Caveat: the press release gives neither a calendar nor a published basis for the tripling of its revenue it announces. Source: https://ocient.com/newsroom/press-releases/ocient-raises-100m-in-series-c-funding/

Anthropic and its allies spent 63 million dollars electing lawmakers

A public policy group backed by Anthropic spent more than 63 million dollars in the US mid-term elections of 3 November, and ten of its seventeen candidates won. It is the first time an AI laboratory has financed elections on that scale. The rival group, backed by an OpenAI co-founder, committed about 54 million dollars. The reversal lies there: elected officials from the open-source camp now defend stricter laws, including one aimed at conversational agents. Source: https://publicfirstaction.us/news/pro-safeguards-candidates-win-in-both-parts-as-voters-reject-big-techs-bid-to-buy-congress

Companies with no real office won approval to buy H200 chips in China

The New York Times visited the addresses of several Chinese companies that Washington authorised to buy or assemble Nvidia H200 chips. In both cases it visited, the offices were empty. One is linked to a sanctioned American investment fund; the other shares staff and patents with Inspur, which the United States lists among entities tied to the Chinese army. The investigation does not say the chips arrived: it says the check relies on paperwork, never on verifying the premises. Source: https://www.nytimes.com/2026/11/07/business/economy/nvidia-chips-chinese-firms.html

Perplexity launches agents that keep working while you sleep

Perplexity launches permanent agents: you give them a goal, they work continuously, hand the work over to each other and alert you when they need a decision. Each one runs on its own virtual machine, with its own screen and browser, and stays reachable from Slack or the app. The service is reserved for subscribers of the Max plan, at 200 dollars a month, and free during a promotional period before usage billing starts. The publisher already had a one-shot tool; this one targets projects that last for days. Source: https://www.perplexity.ai/hub/blog/teammates-are-agents-for-complex-ongoing-work

📊 Trend

Two figures face each other this morning. On one side, 80 % of all the text handled by the main model platform now comes from agents, and corporate AI spending grew 5.4 times in a year: agents have become the first place where AI is actually used. On the other, Google measures that an old comment forgotten in the code can cost an agent 77 % more processed text when it goes hunting a bug that is already fixed. In between sits a legal decision: the licence of a model that fits in a gigabyte forbids commercial use above 10 million dollars of turnover. Scores decide less and less; licences, scope and the measurement of our own agents decide more and more. Our budget should therefore follow the volume actually consumed, not the number of calls. And the question of the month is no longer which model is best, but how far we let our agents act alone. On 12 November, Anthropic's usage policy makes rules about agents connected to physical hardware enforceable. That is a calendar, not a threat.