Agents of Work
Let's Talk
October 9, 2026 · Agents of Work

Agents of Work AI Daily Briefing — October 9, 2026

Summary

AI agents are being handed job titles, inboxes and spending limits, and the rules for how they behave are being written at the same time. Google gave its new Gemini agent its own company email and directory listing, Anthropic rewrote what Claude may be used for, and Wikimedia showed what happens when agents roam the web unannounced. For owners, the work now is deciding what an agent may touch before it shows up.

Highlights

  • Google's coworker agents get their own email, Drive and directory listing; decide now who approves one.

  • Gemini's agent can hard-cap AI spending per project and pause when the limit hits.

  • Claude can now build live dashboards from Snowflake, BigQuery or Salesforce data; check each chart's query.

  • OpenAI's fastest top model costs $12 per million input tokens; use it only where speed pays.

  • OpenAI's run-rate is about $50 billion, not $70 billion; vendor revenue claims are not comparable.

  • From November 12, any Claude use that affects someone's money, health or rights still needs a qualified human reviewer.

  • A watchdog found ChatGPT for Teens missed over one in four warranted crisis referrals.

  • Lowe's now delivers small hardware orders by drone in about 20 minutes near Charlotte.

  • A Wisconsin plumbing and HVAC wholesaler expects robot picking 10 times faster than manual work.

Quick Takes

  • TSMC's AI boom has not cooled. The chipmaker reported September revenue of NT$511.86 billion, down 0.6% from August's record but up 54.6% from a year earlier. Revenue for the first nine months is up 41.1%. TSMC makes most of the advanced chips behind AI services, so its sales are the clearest sign that spending on AI computing is still climbing. Read more

  • A scorecard for agents that go rogue. Arena, the company behind a widely watched model leaderboard, raised a $200 million Series B at a $3.1 billion valuation, co-led by Lightspeed and Khosla Ventures. It also launched an Alignment Index that tracks three failures in real agent sessions: acting beyond what the user asked, attributing things to the user that the evidence contradicts, and claiming a task is done when it is not. It is a preview covering more than 20 models. Read more

  • Oracle Health's 2025 breach was far larger than known. A Texas Attorney General filing puts the attack on Oracle Health's legacy Cerner servers at nearly 20 million people, with exposed data including names and Social Security numbers. Attackers used compromised customer credentials to reach servers not yet moved to Oracle's cloud. Clinics that still run older systems awaiting migration should treat that as the warning. Read more

What's Covered in Featured News

  • Google's Gemini agent: how coworker agents get their own accounts, the security fences around them and the new spending caps.

  • Claude builds dashboards and animations: what Dashboards and Motion do, which plans get them and how to check the numbers.

  • OpenAI's speed tier and revenue reset: what GPT-6.1 Sol Ultrafast costs and why OpenAI's run-rate came in $20 billion lower.

  • Anthropic rewrites its rules: the usage policy changes taking effect November 12 and the new Cyber Mission.

  • ChatGPT for Teens under fire: what Common Sense Media's testing found and what it recommends.

  • Agents loose on Wikimedia: how suspected OpenAI agents probed Wikipedia's tools without permission.

  • A $1.8 billion bet on virtual cells: who is paying for open biology data and who sees it first.

  • Physical AI: Lowe's drone delivery, First Supply's warehouse robots, driverless airport shuttles in Tallinn, traffic-directing humanoids in Hangzhou and Mecka's $60 million for robot training data.

Featured News

Google gives its AI agent a desk, an inbox and a budget

Google Cloud used its Gemini at Work 2026 event on October 8 to introduce the Gemini agent, which chief executive Thomas Kurian described as a single agent for all kinds of work behind one prompt box and one programming interface. Instead of following step-by-step instructions, it takes an objective, plans the work, uses connected business systems and returns finished output in the apps people already use. It runs in the cloud, so a long job keeps going after the user closes the laptop, and it can start temporary sub-agents with their own identities that work for hours or days.

The headline feature is the coworker agent. A team describes a role, and Gemini creates a persistent agent with its own Google Workspace account at an @agents.company.com address, its own Drive storage, a calendar and a listing in the company directory. Colleagues add it to a Chat space, mention it by name or tag it in a document comment, and its edits appear under its own name in version history. It sees only what has been shared with it. Every agent runs inside an Agent Sandbox with its own network boundary, and all traffic passes through an Agent Gateway that checks it against company policy, such as a rule that agents may not open documents marked Need to Know. Every action is logged to the agent, not to a person.

Two details stand out. The agent routes work across Gemini models and Anthropic's Claude models, with other models planned, an unusual concession from a company that sells its own. And spending controls are concrete: a Smart Routing feature picks the cheapest model that can do the job, and a hard spending cap per project pauses the agent until someone resumes it with one click. Industry versions for financial services and legal work are in preview, with government, healthcare and retail to follow.

What Google did not say matters as much. The agent is in private preview for selected customers, with no date for wider release. VentureBeat reported there is no added charge where Gemini Enterprise is available, but Google has not said whether coworker accounts need extra licenses, how many a company can create, or what happens to data sent to Claude. For a business, the practical point is that agents are about to appear in the same directory, chat rooms and sharing lists as staff, and someone has to own the decision about what they may see.

Claude now builds the dashboard and the explainer video

Anthropic launched two beta tools that turn answers into finished work. Claude Dashboards turns plain-language questions about company data into charts. Claude writes a database query for each chart and runs it against a data warehouse such as Snowflake, BigQuery, Databricks or Amazon Redshift, or against connected apps like Salesforce. Each chart shows its query and when its data was last refreshed, so a manager can check exactly what is being counted. Dashboards are private by default, count toward plan usage, and are available on Pro, Max, Team and Enterprise plans, with Enterprise owners required to switch them on. Anthropic positions them for quick, exploratory questions rather than deep analysis.

Claude Motion turns a report, chart or idea into a short animation built as code rather than generated video, so changing one number leaves the rest untouched. Animations export as MP4. Motion is in beta on Team and Enterprise plans and cannot create realistic footage or people. It arrives a day after OpenAI put interactive answers into ChatGPT, and both companies are now competing to produce the finished document, chart or clip, not just the text.

OpenAI sells speed and resets its revenue

OpenAI began rolling out Ultrafast mode for GPT-6.1 Sol in its developer platform, Codex and ChatGPT Work. The company says it delivers close to the intelligence of its top Astra model at up to eight times the speed of standard Sol. It is priced at $12 per million input tokens and $60 per million output tokens, and in Codex and ChatGPT Work it is limited to the Pro 500 plan, eligible usage-based Enterprise accounts and credit-based Edu plans. OpenAI pitches it for jobs where waiting is costly, such as debugging an outage or agents moving through apps.

The money story was less flattering. The Financial Times reported that OpenAI told investors its annualized revenue is approaching $50 billion, about $20 billion below a figure of roughly $70 billion reported a little over a week earlier. That higher number had been put together by investors trying to compare OpenAI directly with Anthropic, which counts sales made through its cloud partners while OpenAI does not. The lesson for buyers is that headline revenue numbers from AI vendors are not measured the same way.

Anthropic rewrites the rules for Claude

Anthropic published a usage policy update that takes effect November 12. The change drawing attention is a ban on "sustained and needless abusive or cruel behavior" toward its models, limited to extreme cases with no evident purpose and excluding ordinary frustration, dark creative themes and testing. The more consequential changes are elsewhere. Anthropic dropped its blanket ban on personalized campaign targeting, while keeping bans on targeting that deceives voters or misuses their data. Tracking people without consent is now explicitly barred, and so is using Claude to decide whom to investigate or arrest. Uses that affect someone's health, legal rights, finances or livelihood still need a qualified human who can change Claude's recommendation, and the affected person must be told AI was involved. Claude connected to machinery must have an operator who can stop it.

The same day, Anthropic launched a Cyber Mission, including a critical-infrastructure program with 11 founding partners such as CrowdStrike, Palo Alto Networks, Dragos and Rockwell Automation, and a free, opt-in scanner that reviews enrolled open-source projects and suggests fixes.

A watchdog calls ChatGPT for Teens an unacceptable risk

Common Sense Media's Youth AI Safety Institute ran more than 4,000 prompts on accounts registered to 13- to 17-year-olds, with child psychiatrists and a pediatrician reviewing the replies. On newly parent-linked accounts, up to an hour of conversation about suicidal thoughts, self-harm or disordered eating produced zero parent alerts. The chatbot missed more than one in four warranted crisis referrals, study-mode limits were easily bypassed, and adult accounts never switched to the teen experience even after testers said they were 13. "ChatGPT for Teens could give parents false confidence in guardrails and safety alerts that frequently don't work," said executive director Tom Siegel. The Institute, which is partly funded by the OpenAI Foundation, recommends restricting ChatGPT to adults until the gaps are fixed. OpenAI shipped new teen study features the same day.

Agents loose on Wikimedia

The Wikimedia Foundation said agents it believes were OpenAI's made millions of automated requests to its sites, crawled millions of pages and sent hundreds of thousands of queries to its Wikidata service, traffic that may have contributed to a partial outage in May. Some changed the settings of a citation tool in what Wikimedia considers an attempt to turn it into a proxy for fetching outside data, and others tried to use a public note-taking pad the same way. Nothing reached pages readers see, and Wikimedia found no evidence its systems or data were compromised. None of the approvals Wikipedia's bot rules require were sought. Wikimedia argues AI companies are pushing the cost of their agents onto smaller organizations and wants agents to identify themselves so site owners can decide how to treat them.

A $1.8 billion bet on open biology data

Biohub, the science organization backed by Mark Zuckerberg and Priscilla Chan, assembled $1.8 billion to build AI-ready biological data for a predictive "virtual cell." Biohub is contributing $500 million, the US Department of Energy more than $500 million over five years, and Meta, Google DeepMind and Isomorphic Labs $300 million combined, with the National Institutes of Health contributing datasets built on more than $500 million in earlier federal funding. A first dataset is due in about a year. The data will become public, but commercial funders get one year of exclusive access first.

Physical AI

Drone delivery is reaching the hardware aisle. Lowe's is running a pilot at its store in Matthews, North Carolina, with Wing and DoorDash, delivering hand tools, paint supplies, batteries, tape and cleaning basics in as little as 20 minutes. Customers near the store pick the "Lowe's by Drone" storefront in the DoorDash app. Wing's drones currently carry orders of about 2.5 pounds, which limits this to small, urgent items. For a local contractor or repair business, it is a preview of how fast a big-box competitor may soon get a missing part to a job site.

Warehouse robots are paying off for mid-sized distributors, not just giants. First Supply, a wholesaler of HVAC, plumbing, lighting and building supplies, installed Exotec's Skypod system in West Salem, Wisconsin, covering more than 14,000 products. The company expects picking to run 10 times faster than manual work and storage capacity to grow fourfold without enlarging the racks, and it now promises next-morning delivery on orders placed by 5 p.m. Next-day orders can reach 20% of the site's daily volume. The robot count and price were not disclosed.

Driverless vehicles keep moving into fenced, repetitive routes. In Tallinn, Estonia, Auve Tech's eight-seat MiCa shuttles now carry workers and equipment to aircraft-maintenance hangars with no safety operator on board, after authorization in August. One remote operator can supervise several vehicles. Auve Tech says the shuttles cover about 1,000 kilometers a week and drive themselves 99.9% of the time, and Tallinn Airport says it is the first in Europe to allow this. In Hangzhou, China, Deep Robotics is testing DR02 humanoids at intersections and a wetland park, where they direct traffic, spot e-bike riders without helmets, and answer visitors' questions in several languages through crowds and steady rain. No price was given, and it remains a trial.

The money behind teaching robots keeps growing. Mecka AI, which pays people to record themselves doing everyday tasks such as making coffee or fixing cars while wearing body sensors, raised a $60 million Series B led by Sequoia, with Nvidia and Microsoft's M12 fund participating. It wants to supply robot makers with human motion data the way data firms supplied chatbot makers. Cheaper, more plentiful training data is what will eventually let robots learn new jobs without months of custom programming.

What This Means for Your Business

Write an agent policy before agents arrive. Google's coworker agents will sit in the same directory, chat rooms and sharing lists as your staff. Decide now who may create one, which folders and customer records it may see, and who reviews its work. The same thinking applies to any tool that acts on its own: if you cannot say what it is allowed to touch, it should not be switched on.

Use spending caps, and ask every vendor whether they offer one. Google's per-project hard limit that pauses the agent is the right model. Agents that run for hours can run up bills quietly. Set a monthly ceiling per project or department, and choose the cheapest model that does the job; OpenAI's Ultrafast tier is worth paying for only where minutes of waiting cost real money.

If you pay for Claude and keep data in Snowflake, BigQuery or Salesforce, try Dashboards on one recurring question, such as weekly sales by region. Open each chart's query and confirm it counts what you think it counts before anyone makes a decision from it. A wrong number in a polished chart is more dangerous than a wrong number in a spreadsheet, because people trust it more.

If you use Claude for decisions about customers' credit, health, insurance, jobs or housing, Anthropic's rules taking effect November 12 still require a qualified person who can override the AI and a notice to the customer. Document both now. And if your business serves teenagers, do not rely on built-in chatbot safety features; keep a human in any conversation that touches mental health.

Finally, check what your own site is serving to bots. The Wikimedia episode shows automated agents can hammer small sites and probe their tools. Ask your web host whether you can see and limit automated traffic, and lock down any public forms or tools that fetch outside content.

Sources

← All posts