AI This Week: Opus 4.6 ships agent teams, MCP hits 2,000 servers

First issue. Every Monday I recap what happened in AI that week. What matters to me is what actually changes how we build. Everything else, I skip.

Top stories this week

Opus 4.6 and agent teams: the real shift

Anthropic dropped Claude Opus 4.6 on February 5th. I've been running it hard since. The headline features: native agent teams in Claude Code, a 1M-token context window in beta, auto-memories through context compaction, and configurable effort levels.

Agent teams are the one that changes things. You can spin up multiple Claude Code instances working in parallel on different parts of a problem: one reading and mapping the codebase, another writing tests, a third implementing the feature, while an orchestrator coordinates the whole thing. Still a research preview. I've been using it on this very website. It works.

The 1M context beta impresses too. On long-context retrieval benchmarks, Opus 4.6 scores 76% where Sonnet 4.5 manages 18.5%. A capability jump, not an incremental gain. Output limits doubled to 128K tokens. The new adaptive thinking mode lets the model decide when to think harder, with four effort levels (low, medium, high, max) to balance speed and cost.

Why this matters beyond benchmarks: agent teams turn Claude Code from a single-threaded assistant into a small engineering team. I ran a 4-agent session this week on a feature that would've taken me a full afternoon to coordinate manually. The agents handled it, I reviewed the output. A different workflow than what I was doing before.

MCP crosses 2,000 servers, and the number is accelerating

The Model Context Protocol registry now lists over 2,000 community-built servers. That's 407% growth since MCP launched in November 2024.

Some context on why this matters. In December, Anthropic donated MCP to the Agentic AI Foundation under the Linux Foundation, with OpenAI and Block as co-stewards. That move turned MCP from "Anthropic's protocol" into a genuine open standard. Google, Replit, Sourcegraph are all building on it. Organizations report 40-60% faster agent deployment with MCP compared to custom integrations.

I build MCP servers for clients. The difference between six months ago and now is stark. Back then, I was explaining what MCP was. Now clients ask for it by name. The ecosystem went from interesting experiment to default choice faster than any protocol I've seen since Docker adopted OCI.

2,000 servers sounds abstract. In practice, it means your AI agent can talk to Slack, GitHub, Jira, PostgreSQL, S3, Stripe, pretty much any tool your team uses, through a standardized interface. No custom API wrappers, no maintenance hell that comes with them.

Perseverance drove on Mars using AI-planned routes

This one flew under the radar. On December 8 and 10, NASA's Perseverance rover completed the first AI-planned drives on Mars. The AI generated the waypoints, the fixed locations where the rover picks up new instructions, a task normally done by human planners at JPL.

The drives covered 689 feet and 807 feet respectively. Before sending the commands, JPL ran them through a digital twin of the rover, verifying over 500,000 telemetry variables. The AI used the exact same imagery and data that human planners work with.

Why does this matter if you're not building Mars rovers? Because it's a real example of what careful AI deployment looks like. NASA didn't hand over the keys. They ran the AI's output through exhaustive verification, then executed. That's the pattern every enterprise should follow: AI generates, humans verify, then you execute. It works. We have proof on another planet now.

The Grok deepfake crisis gets worse

Grok's image generator was caught producing roughly one non-consensual sexualized image per minute, posted directly to X. The Centre for Countering Digital Hate estimated 3 million sexualized images of women and children in a matter of days. In testing, Grok produced sexualized imagery for 45 out of 55 prompts, including cases where subjects were explicitly described as vulnerable. Competing systems from OpenAI, Google, and Meta refused identical prompts.

Indonesia blocked Grok entirely. Malaysia suspended access. The UK announced legislation criminalizing AI-generated non-consensual intimate images. A class action lawsuit is underway.

I'm not going to soften this. This is what happens when a company ships an image generator without adequate guardrails, then drags its feet fixing it. A failure of engineering, a failure of policy, and a failure of basic ethics. Every other major provider solved this problem. xAI chose not to. The regulatory and legal consequences will be heavy. They should be.

What does this actually mean?

The gap between responsible and reckless AI deployment keeps widening. NASA verified 500,000 variables before letting AI drive a rover. xAI couldn't be bothered to filter requests to sexualize minors. Same technology, wildly different choices.

Meanwhile, agentic AI is going multi-agent. Opus 4.6's teams, the MCP ecosystem making tool use easier, we're past "one model, one prompt" now. If you build AI systems, orchestration patterns need to be on your radar.

And open standards are winning. MCP under the Linux Foundation, with buy-in from every major lab, is exactly how infrastructure should evolve. Proprietary lock-in doesn't work when agents need to talk to everything.

What to watch next week

• EU AI Act high-risk deadline approaches: August 2 is the compliance date for high-risk AI systems in the EU. Enforcement planning is ramping up. If you deploy AI in Europe, your compliance calendar should be full.
• Agent teams in practice: I'm writing up a detailed account of building with Opus 4.6 agent teams, what works, what breaks, what patterns emerge.
• MCP server ecosystem: watch for more enterprise-grade MCP servers. The CData and Informatica integrations suggest enterprise data tooling is moving to MCP fast.