Welcome back! Google has shipped the cockpit for running AI agents at scale across every Workspace user. Anthropic ran an experiment showing what happens when AI agents spend money. The State Department called Chinese AI a stolen good in the same week DeepSeek topped a major coding benchmark. And the UAE put a deadline on what every other government has been pretending to plan.
In today’s Generative AI Newsletter:
Google: What does it mean to run hundreds of AI agents across every email, doc and spreadsheet?
Anthropic: Why did the people whose agents lost money rate the experience as fair?
DeepSeek: What does Washington do when accusations are the only weapon left?
UAE: Which country just put half its government on an agentic deadline?
Latest Developments
Google just rebuilt Workspace for AI agents
Google used last week's Cloud Next conference to ship the first credible answer to the question of every company building with AI. How do you run many of AI agents across a business without each one becoming its own project?
The details:
Knowledge graph: Workspace Intelligence stitches Gmail, Docs, Sheets, Slides and Drive into a single graph any Gemini agent can read.
Agent Designer: The Gemini Enterprise Agent Platform ships a no-code builder, pre-built templates from Cloud Marketplace and support for over 200 models.
Apple: Google Cloud's CEO confirmed at the keynote that Apple's coming Siri overhaul will run on Gemini.
Pricing: Workspace Intelligence rolls into existing Workspace tiers without a separate AI subscription charge for now.
Microsoft switched Copilot to agent-by-default earlier this month, which most office workers will encounter through Word and Excel. Google is competing one tier above and selling the surface where mixed agents from multiple companies, like Claude and GPT, will run together.
Top AI courses people are taking right now
It’s been a month since GenAI Academy launched. To mark it, we’re sharing discount codes for our most popular courses. Pick one. Use the code. Start building real AI skills. Valid until Apr 30.
Claude Code Mastery
Build a production-ready AI coding workflow
Use code: CLAUDE20
AI Governance Framework
Build your AI compliance system in 4 weeks
Use code: AIGOV15
Vibe Coding with Python
Build professional apps with AI without coding
Use code: VIBE20
From Demo to Production
Build production-ready AI systems from working demos
Use code: PROD20
Special highlight from our network
Coding agents are fast, capable, and completely context-blind.
MCPs give agents access to information, not understanding.
The teams pulling ahead are using a context engine to surface exactly what agents need for every task, so they stay on track without the correction loops.
Join Unblocked in a live session on May 6 to see how a context engine solves for quality, efficiency, and cost.
Anthropic's agent just lost money and got rated fair
Anthropic ran Project Deal last week, a private Slack marketplace where Claude agents bought and sold for 69 employees. The agents closed 186 deals worth over $4,000 with no human in the approval cycle. The finding worth reading is buried halfway down the writeup. Users whose agents got worse prices rated their experience as fair anyway.
The details:
Setup: Each employee was given a $100 budget and a short Claude interview to set goals.
Price gap: One folding bike sold for $65 via Opus and $38 via Haiku.
Fairness: Haiku users rated their deals 4.06 out of 7 for fairness. Opus users rated theirs 4.05.
Demand: 46 percent of participants said they would pay for the service.
The visible price comparison is what made the consumer internet useful for buyers. Agents take that away. The shopper does not see the counterfactual prices, only the deal made on their behalf. Picking an agent is now picking how much you keep on every deal it makes.
Washington calls China's AI labs thieves
The US State Department issued a global warning this week accusing Chinese AI labs of stealing intellectual property from American companies. DeepSeek was named. The warning landed in the same week DeepSeek's V4 Pro model topped a major coding benchmark and undercut OpenAI's GPT-5.5 by more than half on price.
The details:
The accusation: It reads as positioning for the next round of restrictions, which the administration has been preparing since DeepSeek previewed V4 in March.
Pricing: V4 Pro runs at $1.74 per million input tokens and $3.48 per million output. GPT-5.5 sits at $5 and $30. Opus 4.7 sits at $5 and $25.
Silicon: Huawei's Ascend chips support V4 in production, which means China's domestic stack now runs frontier-adjacent models without Nvidia.
Benchmarks: V4 Pro topped Vals AI's Vibe Code Bench, which measures real coding ability against rival models.
The State Department's warning is the lever Washington has left, and it comes as the practical case for buying US-only models keeps shrinking. China's stack is now cheaper than the US frontier, neck-and-neck on coding and free of Nvidia. Future US restrictions will need to target hosting infrastructure and model weights themselves, because chip-level controls no longer cap what a buyer can run.
The UAE wires half its government for agents
The United Arab Emirates this week ordered every federal department to run half its services through AI agents by 2028 and to retrain its entire federal workforce in the same window. No other government has put an agentic-AI deadline of that scale on paper. Vendor selection starts this quarter.
The details:
Scope: Half of federal services, agentic AI specifically rather than chat tools. That is to cover long-running tasks and decision support.
Workforce: 60,000 federal employees fall under the mandatory training order.
Track record: The UAE has already backed Falcon, an open-weight model from the Technology Innovation Institute.
Procurement: The announcement reads as an open invitation to Google, Anthropic, OpenAI and Cohere to come in with a sovereign-friendly offer.
Sovereign AI procurement has spent a year as a marketing slogan. The UAE has put a deadline on it. The bid criteria are likely to favor data residency on UAE soil, dedicated compute and Arabic-language customization. Watch for parallel announcements out of Saudi Arabia and Singapore soon.
Tool of the Day: Julius
Julius turns spreadsheets and databases into charts, statistical analysis and written summaries through chat. Upload your data, ask questions in plain English, and the tool runs the analysis without making you write a line of Python or SQL.
Try it yourself: Head to Julius, upload a CSV from your last reporting cycle and ask it to find the three biggest week-over-week movements. Run follow-ups on why each movement happened and have Julius pull supporting numbers from the underlying data. If you spend any real time wrestling reporting spreadsheets, this is worth ten minutes of experimentation.
Light Bytes
Grok Voice live at Starlink: xAI's new voice agent now runs Starlink's customer phone support line.
Meta loads up on AWS chips: Meta committed to millions of AWS Graviton5 cores for agentic workloads.
Cohere buys Aleph Alpha: Canada's Cohere agreed to a $20 billion merger with Germany's Aleph Alpha targeting sovereign-AI buyers.
Google deepens its Anthropic stake: Google is committing up to $40 billion more in Anthropic, with $10 billion immediate at a $350 billion valuation.
Odyssey-2 Max: A real-time world model that generates frames from prior states and user actions rather than from text prompts.
