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

Daily Briefing

4 November 2026 · 5 stories

🔥 Top story

1

Anthropic can now prove which model really answered you

When you ask an online assistant a question, you cannot check who actually replied: a middleman can slip in a cheaper model, or someone can edit the text on the way. Anthropic publishes a free prototype, FlashProve, that attaches a mathematical proof to every answer, without ever revealing the inner workings of the model. Tested on Llama 4 Maverick, it produces about 110 proof tokens per second on a single card, and the proof checks out in 17 seconds on an ordinary computer. The price is steep: roughly 70 to 95 times more computing than a normal answer. And the proof only says the right model did the talking, not that the answer is right. Source: Anthropic.

2

CrowdStrike buys two experts in watching and testing agents

When a regulator asks a company what one of its agents did, saying only that it was the model no longer works. CrowdStrike signed two acquisitions on the same day: Fiddler, which watches and explains how agents behave, and Dreadnode, which runs security tests on artificial intelligence automatically. Both join the vendor platform under the names Falcon Observe and Falcon Guardian, with general availability promised in the fourth quarter of its 2027 fiscal year. Together the two deals are valued at 495 million dollars by a Wall Street Journal investigation, an amount CrowdStrike did not confirm and that appears in neither of its announcements. Source: the CrowdStrike press releases.

3

Uber puts one single door in front of 5,000 agent tools

Every team in a big company that wants to plug an agent into an internal service ends up writing its own little connector, with its own security rules. Eventually nobody knows how many tools exist. Uber fixed the problem from the top: one catalogue of more than 800 servers and over 5,000 tools, where an agent looks for what it needs instead of receiving everything at once. The detail that matters most: tools are discovered automatically, but none of them works until the team that owns it has signed off. Uber published its design in October; it is Monday's analysis that brings it back into the news. Source: the Uber engineering blog.

4

OpenAI had its models design an entire device

Two OpenAI engineers designed a complete electronic device in under a week, on the side of their day jobs: one model wrote the requirements, another drew the shape, the case, the circuit board and the software that runs the thing. A manufacturer then produced a first run of prototypes in under a month. The company admits it took several rounds of review before the object was ready to be made. The telling number is elsewhere: since February, the use of these models in the hardware teams has been multiplied by 23. Source: OpenAI.

5

A public benchmark says what agents are worth at fixing bugs

Collinear AI has published the second version of its benchmark, in the open: an agent has to find and fix a security flaw in a real code repository without breaking what worked. Version 2 counts 100 public tasks covering 91 different kinds of flaws. The result deserves a second look: the best model gets only 67.8 per cent of them right, and 19 per cent of the set stays unsolved even for it, even after repeated attempts. In other words, repairing security holes remains hard work for an agent, and its mistakes can go unnoticed. Source: Collinear AI on Hugging Face.

📡 To watch

Honeycomb puts agents on your outages, day and night

The observability company Honeycomb launches agents that watch production releases, catch regressions before customers do and write up incidents. Every finding is attached to the trace it came from and labelled either confirmed or hypothesis. Access is in an early phase. Honeycomb cites a survey in which 76 per cent of teams say code written by artificial intelligence grows faster than their ability to review it, and 65 per cent had an incident caused by an agent loop or a runaway token bill. Source: Honeycomb.

VMware wants agents to query your data without moving it

With its Tanzu platform, VMware offers a way to give agents access to company data without moving everything into one central warehouse. Agents go through a catalogue of governed data products, with exactly the same masking rules as those applied to human readers. The vendor promises fewer tokens burned and predictable cloud exit fees. Source: VMware Tanzu.

Ebix and intentyx watch what your agents really do

The two companies launch a platform that discovers the agents of a company, ties their actions to a declared intent, namely the goal, the authorised tools and the reachable data, and warns when an agent steps outside that frame. The chief executive of Ebix bets that agent governance becomes a security category of its own in 2027. Source: the Ebix press release.

Artefact and Dust rolled out 1,500 agents in a public bank

The French public investment bank Bpifrance ran a four-month trial with one hundred employees, spotted more than sixty possible uses, then moved to five hundred licences and fifteen hundred agents within months. The rulebook, who owns which licence, how risky each agent is and how connectors are watched, was put in place in under four weeks. Source: Artefact.

📊 Trend

One thread runs through today's news: the question is no longer only what a model knows, but what an agent is worth and how you prove it is the one it claims to be. Anthropic invents the mathematical proof. CrowdStrike buys the tools that explain what agents did. Uber locks tool activation behind human approval. A public benchmark says that fixing security holes stays hard work, with 67.8 per cent as the best score. And OpenAI has its models design a whole device. The centre of gravity has moved from the model to the proof, the control and the measurement.