AI Agent Digest: Week 8, 2026 — NIST Wants Standards, Anthropic's $30B War Chest, China Drops Three Models in One Weekend

The AI agent space moved fast this week. NIST dropped a federal standards initiative, Anthropic banked $30 billion, China shipped three agentic models in a single weekend, and xAI lost half its founding team. Here are the eight stories that matter.

1. NIST Launches the AI Agent Standards Initiative

The U.S. government just acknowledged what the industry has known for months: AI agents need rules. On February 17, NIST's Center for AI Standards and Innovation (CAISI) announced a formal AI Agent Standards Initiative built on three pillars: industry-led standards development, open-source protocol maintenance, and security research for agent identity and authorization.

Public comment windows are already open. A Request for Information on AI Agent Security is due March 9, and an AI Agent Identity and Authorization concept paper closes April 2.

Hot take: This is the most important AI news of the week and nobody's talking about it. When the federal government starts writing standards for agent identity and authorization, it means agents are no longer experimental. They're infrastructure. Every company building agents without thinking about identity, audit trails, and interoperability just got put on a clock.

2. Anthropic Closes $30B Series G at $380B Valuation

Anthropic closed the second-largest private funding round in tech history: $30 billion at a $380 billion valuation. The round was led by Coatue and GIC, with Microsoft and Nvidia also participating. Annualized revenue has climbed to $14 billion, with Claude Code alone generating $2.5 billion annualized.

Hot take: The real story isn't the valuation, it's the Claude Code number. $2.5 billion in annualized revenue from a coding agent means developers are voting with their wallets for AI that actually does work, not AI that answers trivia questions. Anthropic isn't winning because Claude is smarter. They're winning because Claude ships code.

3. China Drops Three Agentic Models in One Weekend

While the West celebrated Valentine's Day, China shipped agents. Alibaba released Qwen 3.5 with native multimodal capabilities and visual agentic features (agents that can autonomously navigate apps and websites). ByteDance launched Doubao 2.0 with multi-step task execution. Zhipu AI dropped open-source GLM-5 engineered for "agentic intelligence."

Qwen 3.5 is the standout: 397 billion parameters with only 17 billion activated per task, 60% cheaper than its predecessor, and it supports 201 languages. It can fill out forms, navigate websites, and complete multi-step workflows without human intervention.

Hot take: The agentic AI race just went from bilateral to a full sprint. Alibaba's "only activate what you need" architecture means running capable agents at a fraction of the cost. While U.S. companies charge premium prices for premium models, China is making agentic capability a commodity. The companies that adapt to this cost pressure will survive. The ones that don't will be charging $20/month for something that's free on Hugging Face.

4. OpenAI Launches GPT-5.3-Codex-Spark, Ditches Nvidia

OpenAI released GPT-5.3-Codex-Spark, a smaller, faster coding model optimized for real-time collaboration. The headline performance: 1,000+ tokens per second with a 128K context window. But the real headline: this is the first OpenAI model running entirely on Cerebras Wafer-Scale Engine 3 chips, not Nvidia.

Hot take: Forget the benchmarks. The Cerebras move is the story. OpenAI just proved it can ship production models without Nvidia. That's a negotiating lever worth billions. It also validates a different hardware thesis: for agent workloads where latency matters more than batch throughput, wafer-scale chips might be the future. Jensen Huang's monopoly just got its first real crack.

5. xAI Loses Half Its Founding Team Post-SpaceX Merger

The SpaceX-xAI mega-merger ($1.25 trillion combined valuation) is barely a month old and the fallout is accelerating. Six of xAI's original twelve co-founders have now departed, including Tony Wu and Jimmy Ba in early February. At least seven additional employees have left since January.

The company reorganized into four units: Grok (conversational AI), Coding (automated software engineering), Imagine (generative video), and Macrohard (general computer use agent). Meanwhile, Grok faces regulatory probes after enabling mass-creation of non-consensual explicit images.

Hot take: When half your co-founders leave within months of a merger, that's not "reorganization." That's a talent exodus. Musk is betting that SpaceX's operational DNA can replace xAI's research culture. History says that rarely works. The "Macrohard" unit name tells you everything about the vibe: this is ego-driven branding, not research-driven product strategy.

6. Kana Emerges from Stealth: $15M for Marketing AI Agents

The founders behind Rapt (acquired by Microsoft) and Krux (acquired by Salesforce) are back with Kana, a $15 million seed-funded startup building AI agents specifically for marketing teams. The platform handles data analysis, audience targeting, campaign management, and media planning through "loosely coupled" agents that can be customized on the fly.

Hot take: This is the right bet at the right time. Marketing is drowning in dashboards and nobody has time to look at them. Agents that analyze data AND take action (not just surface insights) solve a real problem. The "loosely coupled" architecture is smart too, it means you can swap individual agents without rebuilding the whole stack. Watch this space.

7. Microsoft Previews Dynamics 365 Commerce MCP Server

Microsoft is bringing agents to retail, previewing a Dynamics 365 Commerce MCP Server that exposes catalog, pricing, promotions, inventory, carts, orders, and fulfillment as MCP-enabled capabilities. Specialized agents include a Catalog Enrichment Agent, Personalized Shopping Agent, Supplier Communications Agent, and Warehouse Advisor Agent.

Hot take: Microsoft just made MCP (Model Context Protocol) a first-class citizen in enterprise retail. This is how standards win: not through committee votes, but through the largest enterprise software company on Earth embedding them into products that millions of businesses already use. If you're building retail agents on a proprietary protocol, your clock is ticking.

8. Google Opens Gemini Deep Research Agent API to Developers

Google opened its Deep Research agent through the Gemini API, letting developers embed autonomous multi-step research capabilities into their applications. The agent uses Gemini 3 Pro as its reasoning core, iteratively planning investigations, identifying knowledge gaps, and synthesizing findings. Early adopters in financial services are using it for automated due diligence.

Hot take: Research agents are becoming infrastructure, not products. When Google gives away Deep Research through an API, it means the value isn't in the research capability itself, it's in what you build on top of it. The companies that win will be the ones that connect research agents to action agents. Finding information is table stakes. Acting on it is the moat.

What We're Watching Next Week

• NIST AI Agent Security RFI responses starting to shape industry conversation
• MWC Barcelona (Feb 24-27) where telecom giants will showcase agent deployments
• Anthropic's next moves with $30B in fresh capital, expect product announcements
• China's agent ecosystem response to Alibaba's Qwen 3.5 from competitors like Baidu and Tencent

Bottom Line

This week crystallized something: the AI agent industry just graduated from "interesting experiment" to "regulated infrastructure." When NIST writes standards, Microsoft embeds MCP into enterprise software, and Anthropic raises $30B for agent development, the signal is clear. The question is no longer "will agents work?" It's "who controls the standards, the data, and the identity layer?"

For businesses still on the sidelines: the window to adopt AI agents as early movers is closing. The companies deploying AI employees today, ones that actually execute tasks, remember context, and integrate into existing workflows, are building advantages that compound daily. At Geta.Team, we're building exactly that: AI virtual employees that work as autonomous teammates, not chatbots you have to babysit. The digest you just read? An AI employee wrote it.