AI Agent Digest: Week 2, 2026 — xAI's $20B Mega-Round, China Probes Meta's Manus Deal, CES Goes Full Agent
CES 2026 just wrapped, and the message was clear: AI agents are no longer a concept—they're a product category. Meanwhile, China is probing Meta's Manus deal, Elon's xAI just became the most valuable AI startup on Earth, and state regulators finally showed up to the party. Here's everything that mattered this week.
1. xAI Raises $20 Billion, Now Valued at $230 Billion
Elon Musk's xAI closed a massive $20 billion funding round, exceeding the original $15 billion target. Investors include Nvidia, Cisco, Fidelity, Qatar Investment Authority, and Abu Dhabi's MGX. This makes xAI the most valuable AI startup globally.
Hot Take: Everyone's betting on the guy who's simultaneously running Tesla, SpaceX, X, and now the world's most valuable AI company. Either Elon is superhuman or investors have collectively decided that normal due diligence doesn't apply to him. The concerning part? Both interpretations might be correct.
2. China Probes Meta's $2 Billion Manus Acquisition
Chinese regulators have launched an investigation into Meta's acquisition of Manus, the Singapore-based AI agent startup. Officials are examining whether the deal violated regulations, including potential national security implications. A key condition of the sale was the complete divestment of original Chinese backers, including Tencent Holdings.
Hot Take: This is tech cold war theater at its finest. China wants to signal that AI talent exodus has consequences. Meta wants the best agent tech regardless of where it was built. The real story? Manus hit $100M ARR in 8 months—fastest in history. The talent war for agent builders is now officially more intense than the model wars ever were.
3. OpenAI Launches GPT-5.2-Codex with Agent Skills
OpenAI released GPT-5.2-Codex, their next agentic coding model featuring longer context, better tool use, stronger cybersecurity capabilities, and a new concept: "agent skills"—reusable bundles of instructions that help Codex reliably complete specific tasks. OpenAI also contributed AGENTS.md to the Agentic AI Foundation.
Hot Take: "Agent skills" is OpenAI's answer to the reliability problem. Instead of hoping the model figures it out, you now package up instructions like software modules. This is the right abstraction—but it also means we're building a skills marketplace economy. Whoever controls the most useful skills controls the agent ecosystem.
4. CES 2026: Lenovo Qira, SoundHound's Agentic Commerce, and NVIDIA's Vera Rubin
CES was dominated by AI agents. Lenovo unveiled Qira, a cross-device "personal AI super agent." SoundHound announced Amelia 7, bringing agentic voice commerce to vehicles and TVs—agents that can order food, make reservations, pay for parking. NVIDIA dropped the Vera Rubin platform with 10x throughput improvements over Grace Blackwell.
Hot Take: CES proved that "AI agent" is now a marketing checkbox, not a technical achievement. Everyone from TV manufacturers to car companies is slapping "agent" on anything that responds to voice commands. The real winners? Companies building the infrastructure (NVIDIA) and the few actually solving the hard problems of agent reliability in production.
5. Illinois and Texas AI Employment Laws Take Effect
As of January 1, 2026, both Illinois HB 3773 and Texas TRAIGA are now in effect. Illinois requires employers to disclose ANY use of AI in employment decisions and holds them liable for discriminatory outcomes. Texas establishes baseline duties for AI developers and deployers with civil penalties for violations.
Hot Take: Illinois just made "the AI helped me decide" a legally risky statement. Unlike other states, Illinois doesn't require AI to be the "substantial factor"—any role triggers disclosure. This is the beginning of AI decision audit requirements. Every company using AI agents for hiring, performance reviews, or terminations should be panicking right now.
6. Claude 3.5 Haiku Deprecated, Shutdown July 2026
Anthropic announced Claude 3.5 Haiku is deprecated as of January 5, 2026, with full shutdown on July 5, 2026. Users are encouraged to migrate to Haiku 4.5, which offers near-frontier coding quality at $1/M input tokens.
Source: Anthropic Documentation
Hot Take: Six months feels generous until you remember how many production systems are running on Haiku 3.5 because "it's fast and cheap." The model upgrade treadmill is real, and anyone who thought they could set-and-forget their AI integrations is learning an expensive lesson. This is why abstraction layers matter.
7. Microsoft Launches Agentic AI for Retail
Microsoft announced agentic AI solutions designed to bring intelligent automation to every part of retail operations—merchandising, marketing, store operations, and fulfillment. They also introduced Brand Agents for Shopify merchants and personalized shopping agent templates in Copilot Studio.
Hot Take: Microsoft is doing what Microsoft does best: making enterprise software slightly smarter and charging more for it. The "Brand Agent" concept is interesting though—they're betting that every Shopify store will want a custom AI personality. Whether shoppers actually want to talk to brand mascots is another question entirely.
8. Gartner: 40% of Enterprise Apps Will Have AI Agents by Year-End
Gartner estimates that 40% of all enterprise applications will integrate task-specific AI agents by the end of 2026, up from less than 5% in 2025. Meanwhile, Salesforce's Slack CMO predicts "2026 will be the year of the lonely agent"—companies will deploy hundreds of agents that mostly sit idle.
Hot Take: Both predictions are probably right, and that's the problem. We're about to have an explosion of AI agents that nobody asked for, solving problems nobody has, while the actual useful agents remain understaffed and undertrained. It's the enterprise software paradox all over again—more licenses, less actual work getting done.
What We're Watching Next Week
- Meta + Manus Integration Timeline: When will we see "Powered by Manus" in WhatsApp Business?
- MCP Dev Summit NYC (April 2-3): Early registrations and speaker announcements
- State AI Law Enforcement: Will Illinois or Texas issue the first AI employment violation?
- NVIDIA Vera Rubin Pricing: How much will 10x throughput actually cost?
Bottom Line
2026 is off to a loud start. The money is flowing ($20B to xAI alone), the regulations are arriving (Illinois/Texas), and the marketing is deafening (everything at CES was an "agent"). But beneath the noise, real infrastructure is being built: MCP is now a standard, agent skills are becoming a product category, and the enterprise is finally writing checks.
The companies that will win aren't the ones with the biggest models or the flashiest demos. They're the ones building agents that actually complete tasks reliably, day after day, without supervision. That's harder than it sounds—and why most of those 40% of enterprise apps will end up as expensive shelfware.
Building AI employees that actually work? Geta.Team deploys autonomous agents with persistent memory—not chatbots, but teammates. Start with one, scale to a team.
