Google hands the office to a single AI agent
Edition of 2026-10-09
# AI For All ### Friday, October 9, 2026
As of today, an operations manager at a large company can hand an outcome to an agent that plans the work, spins up its own helper agents with corporate email addresses, and delivers it inside the documents and inbox the team already uses.
Your next colleague is an agent with its own email address
At its Gemini at Work event on Thursday, Google Cloud launched the Gemini agent, a single assistant built to take an outcome rather than a list of instructions. You assign an objective. It plans the work, uses tools, and returns something finished inside the documents, the inbox, and the developer environment you already use. It can create temporary sub-agents, give each a corporate identity and an email address, and keep working in the cloud after you close your laptop. It also chooses the model for the job, Google's own or Anthropic's Claude, inside the same interface.
Google is selling this to companies first, and it leads with scale. It says nearly 80% of its cloud customers already use its AI products, and that nearly 90% of the Fortune 100 use Gemini Enterprise. The part meant to answer the obvious objection is governance: each agent gets role-based permissions approved by a security administrator, and all of its traffic passes through one gateway that enforces company policy in real time.
The reservation is the one that follows every agent launch. An agent that plans its own steps is also an agent that can take the wrong ones without asking. That is why the permission list, and not the demo, is the thing to read.
A free door opens for the people who fix our software
Anthropic launched the Anthropic Cyber Mission on Thursday, an effort to put frontier models in the hands of defenders rather than attackers. Two pieces land now. The Critical Infrastructure Defense Program works with the companies that protect power grids, water systems and transport networks, and its founding partners include Accenture, CrowdStrike, Deloitte, Palo Alto Networks, PwC and Rockwell Automation. OSS Scanner is free and opt-in: it runs regular security scans on open-source projects and sends the findings straight to the maintainers.
Anthropic is unusually direct about the trade-off. The reports are written by the model and sent with no human review, so they arrive fast but can be wrong, and the company expects a true-positive rate above 90%. It published a new usage policy the same day, so its whole week has been about who gets access to its strongest models.
If you run the code or the equipment that other people depend on, this is the week the door opened. Free, fast and imperfect beats nothing at all.
The best cyber model is the one you cannot have
An independent board that ranks models on defensive security work, Artificial Analysis, added a new category on Thursday. Until now it measured only the models anyone can reach. The new one covers "trusted-access" models, kept behind vetted programs and shipped with fewer safety blocks. The first entry, OpenAI's GPT-6 Sol (Daybreak Blue), now leads the board. It triggers no safety refusals across the index, and it beats the publicly available version of the same model by 32 points.
The price gap is as wide as the capability gap. Artificial Analysis puts it at about $1.77 per task, against about $11.67 per task for Grok 4.7 at its highest effort setting. That is the shape of a two-tier market: the strongest and cheapest tool for the job is, by design, not sold to the public.
For anyone choosing how to defend a company's systems, the useful takeaway is not the ranking but the pattern. Capability is being held back on purpose, and the version you can buy is not the version that wins.
Weak signal
Anthropic rewrote its Usage Policy this week, effective 12 November. It gathers the rules against deceptive activity into one section, sharpens its election rules, and removes a blanket ban on personalized vote and campaign targeting, which had blocked legitimate work such as multilingual voter information. It also bans sustained cruelty toward the model itself. There is nothing to download here, which is why it stays a signal. But a company building on Claude is, today, reading the final draft of the rules it will live under.
Who gets the strongest model, and what they may do with it: both answers moved this week, and neither is settled.