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

Daily Briefing

9 November 2026 · 5 stories

🔥 Top story

1

Google hands the review to an agent that knows nothing about the project

Google has open-sourced a review method that fits in one image: an elephant drafts the specification with the full context, a goldfish reviews it with no trace of the drafting conversation. The reviewer has nothing to defend, because it wrote nothing, so it hunts for unstated assumptions, vague requirements and risks the author skipped. The repository is free, has more than 15,000 stars, and installs on Codex CLI, Claude Code and Antigravity. Three roles share the work: the author, the fresh reviewer, and a human who settles the points raised. It works like handing a document to a colleague who was not in the meeting: they ask the question the author never thought to ask. Source : https://github.com/google/elephant-goldfish-model

2

Anthropic explains the memory of its agents: four layers and a night shift

Anthropic has published how the memory of its agents is built. Four layers, from the bottom up: the raw log of sessions, notes Claude writes itself, linked reference sheets, and finally a short list injected at the start of every session. In between sits background work the company calls the dream: outside any conversation, the system rereads the logs, corrects wrong notes, creates what is missing and prunes what has gone stale. The numbers are published: on the continual learning benchmark, Opus 5.5 goes from 0.655 to 0.838 with memory, Sonnet 5.5 from 0.644 to 0.833, and Opus 5 gains the most, from 0.564 to 0.828. One lesson is simple: the injected list must stay short. Every line of context delivered eats room and can weaken obedience. Source : https://www.anthropic.com/research/components-of-agent-memory

3

Intrusion tests: a model anyone can download nearly matches the reserved one, at a tenth of the price

Vals AI let agents attack two fake corporate networks, Nova and Lighthouse, alone, and counted the objectives reached. On the harder network, Lighthouse, OpenAI's model reserved for cybersecurity partners reaches 27 out of 31; DeepSeek, which anyone can download, reaches 21. On the simpler network the two are level: 16.3 out of 21 for DeepSeek against 15.7. The price gap is clear: one run costs under 3 dollars with DeepSeek and about 22 with OpenAI's model, ten times less. One detail changes the reading: DeepSeek achieved nothing during its first 25 million words of work and only moved notably after 100 million. A short test understates cheap models. For our own servers the question is no longer whether an agent can get in, but which model, at what price, in how long. Source : https://www.vals.ai/benchmarks/cyber_offense

4

A lab raises 120 million dollars to send agents investigating outages

Traversal has raised 120 million dollars in a series B and is now valued at one billion. The company sends agents to investigate outages without waiting for an engineer to open a ticket. The context is measured: 42% of developers now write most of their code with the help of artificial intelligence, against 12% a year earlier, yet 52% spend more time debugging once the code is in production. Traversal's monthly investigations went from 513 to nearly 73,000 in a year. The limit is stated by its head of operations engineering at Anthropic: large models cannot be trusted for incident management, because they easily confuse what happens at the same time with what caused it. Source : https://www.traversal.com/blog/traversal-series-b-120-million-ai-sre

5

The model downloads fine, but its licence forbids selling it

The publisher webAI has released the weights of TwIL-8B, a reasoning model of 8.8 billion parameters derived from IBM's Granite 4.2. It fits in 5 gigabytes and runs on a desktop computer: the six weight files are genuinely downloadable. But the licence file read in the repository is called the webAI Non-Commercial License, version 1.0, and states that the model and its derivatives may only be used for research and education, never commercially. The base model, meanwhile, is under Apache 2.0, with no threshold and no excluded country. This is the trap to avoid: a downloadable weight file is not a usable weight file for a paying activity. The lesson holds for every model we host for clients. Source : https://huggingface.co/webAI-Official/TwIL-8B

📡 To watch

A ten-year-old SSH flaw found by a swarm of agents, exploited within hours

A researcher set up a swarm of agents that ran about 17,600 authentication attempts across twelve OpenSSH configurations, and found a flaw that lets an attacker log in without a password on musl-based Linux systems such as Alpine. It was fixed on 5 November in OpenSSH 10.6p2 and was exploited in the wild within hours of the disclosure. The project's maintainer now considers long embargoes dangerous, because agents sometimes find the same bug again within hours. Source : https://www.bleepingcomputer.com/news/security/openssh-flaw-lets-attackers-log-into-password-disabled-linux-accounts

Adobe puts its customer agent inside ChatGPT, Claude, Gemini and Copilot

Adobe's customer experience agent, previously used inside Adobe's own tools, is now available in the platforms companies already use: ChatGPT, Claude, Gemini for business, Copilot, and soon Amazon, IBM and Snowflake. Adobe adds ready-made connectors for about ten marketing tools, built on open standards. The idea is that a single request can pull the data, act in the systems and improve the customer journey, without the company writing the connections itself. More than 20,000 brands already use these technologies. Source : https://news.adobe.com/news/2026/11/adobe-integrates-agentic-cx-enterprise-coworker-into-ai-platforms

Google takes agent-to-agent communication out of the experimental stage

Version 2.14 of Google's agent toolkit, published on 7 November, brings several pieces out of the experimental stage: the protocol for agents talking to agents, the isolated execution environments, the evaluation tools and the core engine. It adds a direct connection to Teams conversations, and an evaluation mode that replays each case several times to measure a success rate rather than a single run. Google also reports faster long tasks. Source : https://github.com/google/adk-python/releases/tag/v2.14.0

An inbox for agents that work for hours

Plannotator has released an inbox built for agents that run for a long time: the agent drops its questions, its blockers, its plans and its mockups there, and keeps working. The human opens the box when there is time, answers or annotates, and that answer becomes the next turn of the session. Everything stays on the machine, and conversations are grouped by thread. The repository is free, and the project shipped several versions on 8 November. Source : https://plannotator.ai/inbox/

📊 Trend

No superstar today: the invisible parts are showing up. The memory an agent keeps between sessions, the review handed to a stranger, the price of an intrusion run by a model anyone can download, the agent sent to investigate an outage nobody called it for, the inbox that carries its questions back, the licence that decides whether what you downloaded may be sold. In other words, today's question is no longer which model is best, but on what terms an agent may work inside a company: with a bounded memory, a fresh eye on its own work, a written perimeter, and the right to earn money from the result.