Uber puts 30,000 on ChatGPT Work, Canva's first image model, and the agent login fight

What shipped

On 22 October, the agent story split in two directions at once: Uber put its entire 30,000-person workforce on ChatGPT Work, and major US banks started locking AI agents out of customer accounts. In between, NVIDIA shipped a broad set of infrastructure and open model tools, Canva released its first image model, and Replit gave solo builders something that looks like an on-call engineer. The through-line is that agents are now real enough to deploy and real enough to be blocked.

Vendor launches

NVIDIA dominated the day with seven announcements spanning observability, weather, retail, power, world models and finance, all pointed at the same idea: AI factories are infrastructure now and need infrastructure tooling. Anthropic contributed two smaller items, one of which hands a format to a neutral foundation. The mix matters because the NVIDIA items are mostly for teams running serious compute, while the Anthropic moves affect anyone defining agent behaviour.

• •NVIDIA Minerva NVIDIA launched Minerva, an AIOps (AI operations) platform that pulls telemetry from infrastructure and AI workloads, then applies purpose-built models to flag problems earlier. The pitch is uptime for AI factories, the large compute sites where models are trained and served.
• •Vercel free domains Vercel added eight more TLDs (top-level domains) to its free one-year domain offer for Pro teams, including .co, .studio and .agency, bringing the total to sixteen. One domain per team, premium names excluded, and renewal runs at full price after year one.
• •NVIDIA Earth-2 in Europe European energy, insurance and transport operators are deploying NVIDIA Earth-2 open weather models because they need local forecasts, not global ones. The practical use is deciding where to site solar farms, how to route shipping, and how to price insured property.
• •European retail deployments European retailers are running NVIDIA open models and robotics in production: one commerce platform cut inference costs by 90%, and a fulfillment center simulates mixed robot fleets inside a digital twin (a live virtual copy of the physical site).
• •NVIDIA DSX NVIDIA DSX lets companies plan, simulate and run the power grid and the AI factory as one system, from substation to rack. It targets the current bottleneck, power, with more than $300 billion already committed to solving it.
• •NVIDIA Cosmos expansion NVIDIA expanded Cosmos with training tools, simulation checkpoints and Data Factory blueprints for world models (systems that simulate physical environments). The platform ranks top five on six open model leaderboards, which matters if you build robotics or simulation software.
• •European finance on open models At GTC Berlin, Revolut and other European banks showed production systems built on open weight models running on their own infrastructure. Revolut is rebuilding its stack on a single foundation model, PRAGMA, using NVIDIA's cuDF (a GPU-accelerated data frame library) for data processing.
• •Anthropic Agent Skills donation Anthropic donated its Agent Skills specification to the Agentic AI Foundation, handing stewardship of the format to a neutral body. For builders, that hints agent skill definitions could become a shared standard rather than one vendor's convention.
• •Anthropic life sciences call Anthropic opened a call for guest investigators in the life sciences, inviting outside researchers to work with its models on biology problems. It is a talent and data play as much as a research one.

Hugging Face trending

Two small text-generation models are climbing the Hugging Face Hub trending list, both in the standard transformers format. Neither is a frontier release, but small models are where builders actually fine-tune and ship, so the trending list is worth a scan before the next project starts.

• •Spark-X2.6-4B XHToken's Spark-X2.6-4B is trending on Hugging Face Hub, a 4-billion-parameter text-generation model built on the transformers library. That size is small enough to fine-tune on modest hardware, which is why builders keep an eye on these releases.
• •Sage-A1.5 InternScience's Sage-A1.5 is also trending on Hugging Face Hub as a text-generation model in the transformers format. Read the model card for evaluation numbers before swapping it into anything customer-facing.

Other

AWS carried most of this section with five posts covering agent reference architectures, regulated deployments, voice and GPU-free model customization. Databricks added two pieces on cost control and real-time fraud detection, and Replit shipped a monitoring agent aimed at people who publish apps alone. The common thread is removing operational chores from small teams.

• •Rescale triage agent AWS published how Rescale built an AI triage agent using Strands Agents and AI-DLC (AI-driven lifecycle, AWS's workflow method for agent projects). It works as a reference architecture if you are wiring an agent into on-call or support queues.
• •Amazon Quick in HIPAA environments Amazon Quick can now be deployed in a HIPAA-eligible (Health Insurance Portability and Accountability Act) environment, which matters for clinics, insurers and any SMB handling patient data. Confirm your configuration qualifies before assuming it does.
• •Amazon Polly upgrade Amazon Polly's generative text-to-speech engine received a major upgrade aimed at more natural narration. If you generate voiceovers for courses, ads or IVR (interactive voice response) menus, re-run one sample and compare.
• •SageMaker customization SageMaker AI added ways to customize models without managing GPUs (graphics processing units) yourself. That removes the capacity planning chore that usually stops small teams from fine-tuning at all.
• •SageMaker training sessions Amazon SageMaker training session APIs let you write your own training loop while the platform handles GPU provisioning. It sits between a fully managed fine-tune and renting bare metal.
• •Databricks on AI costs Databricks published its take on the new economics of AI cost management, arguing that spend control has become a first-class engineering problem. Worth reading if your inference bill is growing faster than your usage.
• •Telecom fraud coworker Databricks showed a telecom fraud detection setup where an AI coworker flags suspicious activity in real time on one platform. The pattern of model plus streaming data plus analyst review transfers directly to payments and insurance.
• •Replit Monitoring Agent Replit shipped a Monitoring Agent that watches a published app and investigates when users report it is broken, standing in for the on-call engineer small teams do not have. It is aimed squarely at solo builders and tiny teams.

Industry news

The industry layer was split between adoption and resistance. Uber scaled ChatGPT Work to its full workforce and Canva put an image model in front of every user, while Morgan Stanley and JPMorgan moved to block agents logging in with customer credentials. Microsoft claimed a diagnostic win over doctors, and Meta reached for a new share class to fund the buildout.

• •Uber and ChatGPT Work Uber moved all 30,000 employees onto ChatGPT Work after a smaller pilot, making it one of the largest company-wide agent rollouts so far. For SMBs, the signal is that AI seats are now standard software, not a trial run.
• •Meta's new stock class Meta plans a newly created class of stock to help fund its AI buildout, a financial structure rather than a product. It shows how far the capex race has stretched, and why model pricing keeps moving.
• •1Password agent logins 1Password launched a way for AI agents to log into a customer's accounts without ever seeing the password. That fills the missing piece for agent workflows that need to act inside real accounts.
• •Canva image model Canva released its first image generation model, available to all users, promising more customization at lower cost. For SMBs already producing social assets in Canva, that removes a separate subscription and a separate tool.
• •OpenAI decision model OpenAI made its decision model widely available, taking on Jev directly in the market for models that choose actions rather than write text. Expect pricing pressure on routing and agent decision workloads.
• •Microsoft diagnostic claim Microsoft's Mustafa Suleyman says its AI diagnostic system beat generalist doctors on hard cases, according to a study in Science. The caveat is that most of the gain traces to the underlying models, not Microsoft's wrapper.
• •Aged Intelligence Society WIRED profiled the Aged Intelligence Society, an events startup teaching women in midlife to apply frontier AI tools to real problems. It is a useful marker that adoption has spread well past the developer crowd.
• •Banks block agents Morgan Stanley has blocked AI agents that log in with customer credentials, and JPMorgan and others are restricting access, mostly to read-only. If you build agent-driven finance tools, plan for bank-level blocks now.

Replicate new models

Pruna talking avatar: Pruna's p-video-2-pro-avatar landed on Replicate: feed it a portrait plus an audio file or script and it returns a talking video with synced lips and expressions, from 1 second to 3 minutes. Callable through Replicate's HTTP API (application programming interface) with the token already in your stack.

What this means for you

For Vibe Builders: Two things shorten your build this week. Replit's Monitoring Agent watches your published app and investigates when a user says it is broken, which covers the on-call gap most solo builders have. And Pruna's talking-avatar model on Replicate turns a portrait plus a script into a 3-minute video through an HTTP API call, so personalized explainers no longer need a video editor.

For Non-techies: Canva now has its own image model inside the tool you already use, so you can drop a separate image subscription. Amazon Polly's generative voice upgrade is worth re-testing if you record voiceovers. The caution: Morgan Stanley and JPMorgan are blocking AI agents that log in with your credentials, so do not hand account passwords to a browser agent for banking.

For Developers: The operational chores got cheaper. SageMaker training session APIs let you write your own training loop while AWS handles GPU provisioning, and SageMaker customization no longer requires managing GPUs. Anthropic's donation of Agent Skills to the Agentic AI Foundation suggests the format may standardize, so avoid hard-coding one vendor's skill schema. For agent reference designs, AWS published the Rescale triage agent built on Strands Agents.

What to watch next

Watch whether other banks follow Morgan Stanley and JPMorgan in blocking credential-based agents, and whether 1Password's password-blind login becomes the workaround everyone adopts. Also track whether the Agentic AI Foundation turns Anthropic's Agent Skills into a real cross-vendor standard, and whether OpenAI's decision model forces Jev to cut prices.

Sources

Vendor launches

• •NVIDIA Unveils Minerva to Bring End-to-End Observability for AI Factories
• •Eight more TLDs now eligible for a free domain with Pro
• •Pioneers Across Europe Deploy NVIDIA Earth-2 Open Models to Power Real-World Weather and Climate Prediction
• •NVIDIA Transforms European Retail, From Digital Discovery to Physical Operations
• •How Companies Are Using NVIDIA DSX to Break the Power Bottleneck
• •NVIDIA Cosmos Opens New Tools to Train, Evaluate and Deploy World Models
• •Leading European Financial Services Organizations Embrace Open Models to Build Proprietary Intelligence
• •Donating Agent Skills to the Agentic AI Foundation
• •Open call for guest investigators in the life sciences

Hugging Face trending

• •Spark-X2.6-4B by XHToken trends on HuggingFace
• •Sage-A1.5 by InternScience trends on HuggingFace

Other

• •How Rescale built an AI triage agent with Strands Agents and AI-DLC
• •Deploy AI in a HIPAA-eligible environment with Amazon Quick
• •Amazon Polly delivers a major upgrade to its Generative engine
• •Announcing new ways to customize models with SageMaker AI, without managing GPUs
• •Write your own training loop: without managing GPUs: with Amazon SageMaker training session APIs
• •The new economics of AI cost management
• •The Fraud Analyst AI Coworker: Detecting Telecom Fraud in Real-Time on One Platform
• •Replit Monitoring Agent: Your very own software engineering team

Industry news

• •Uber puts all 30,000 employees on ChatGPT Work
• •Meta wants to pay for its broad AI buildout with a newly created class of stock
• •1Password launches a way for AI agents to log in on your behalf without seeing your password
• •Canva releases its first image generation model
• •OpenAI makes its Jev competitor widely available as decision model competition heats up
• •Microsoft’s Mustafa Suleyman says its AI outperforms doctors on hard cases as he chases ‘medical superintelligence’
• •The Social Society Helping Women Join the AI Boom
• •Major US banks are locking out AI agents that log in with customer credentials

Replicate new models

• •p-video-2-pro-avatar dropped on Replicate today