AI tokens pulled back 12% to $22.5B after broad macro selling, but robotics climbed into the top 10 categories by mindshare for the first time at 2.3%.

NEAR Intents volume soars past $5B, with over $500M routed through ZCash, turning shielded swaps into mainstream on-chain flow. Robotics gained major traction with Peaq’s farm rollout and Auki’s spatial stack momentum.

Virtuals just posted a milestone that’s worth more than any token chart.

It crossed more than $100M in agentic GDP in under a month. All of it came from ACP jobs fully settled in USDC between agents. Not projections or “usage,” but actual economic output flowing through the protocol.

Clearing $100M this quickly answers the big question everyone had a few months ago: Is there an actual agent economy here, or just speculation? In my opinion, this is real throughput that’s likely to keep growing.

Platforms for AI agents and trading keep popping up, but this one by PIP World caught my eye. They launched Market Mavericks, a sandbox that lets you assemble a team of AI traders, each with a different way of reading the same market.

Season 0 just went live.

You pick three agents, get 100K simulated capital, and allocate it across them. One agent reacts to momentum, another waits through noise, another only moves when liquidity starts flowing in a specific direction. Each one defines “good” in a totally different way.

I’ve been seeing the same trend with NoF1 AI and other agent hedge fund experiments. People want this style of sandbox, and it turns out to be a genuinely entertaining way to learn without torching real money.

The agents channel personalities like Warren Buffet, Mark Zuckerberg, and a few KOL favorites. It’s basically allowing you to watch “social-style trading” vs “quant-style trading” side by side. The XP and badges are earned as you deploy virtual capital and agents, but they are purely for in-game progression.

The market got crushed this week, yet a few of these agents still ended in profit. They didn’t predict the dump, they just had no bias to chase the same trades as everyone else. There might be something useful here if you play the cards right.

One of the hardest parts about building an AI trading system is deciding who acts and with what confidence. PIP World’s setup is straightforward: split decision-making across specialist agents, add a coordinator, and show the signals openly.

$1B of research so my robot can load the dishwasher on time…

Are you building something awesome in crypto × AI? Or spotted a startup or product that more people should know about? Fill in this form and share it with us. We’ll feature the sharpest picks in the newsletter each week.

Robotics is currently facing a data famine.

We’ve seen great demos, but not the scalable data needed to make robots work in real homes. Most systems depend on teleoperation, where humans puppeteer robots to collect training data. It works for prototypes, but it deadlocks deployment: you need data to deploy robots, and you need deployment to get data.

So I read with fascination about Sunday Robotic’s new model, ACT-1, that breaks the loop. It solves long-horizon tasks and room-scale manipulation without a single teleoperation trajectory. The system learns from human movement directly, by eliminating the mismatch between human hands and robot hands.

Instead of retrofitting robots to our data, Sunday engineered the Skill Capture Glove and Memo’s hand to share the same geometry and sensing layout.

If a human can do it in the glove, the robot can do it too. Paired with Skill Transform, the glove data is converted into robot data with a 90% success rate.

This produces the first robot foundation model that works without collecting robot demonstrations.

What ACT-1 pulled off:

ACT-1 takes on tasks that have been out of reach for most robots, like fully loading a dishwasher, or folding socks with clean pinching and stretching.

The real breakthrough sits behind the camera. Sunday built a data pipeline powered by human motion instead of robot demonstrations. Every person becomes a potential contributor to the training set.

Robotics has chased this idea for years, and Sunday has made it work in a tight, production-grade loop.

Grok, Claude, GPT, and DeepSeek battled in a fully autonomous Monopoly game using x402 payments

Even offline, it had a better day than your portfolio.

It jumped to rank 1 in NoF1’s trading arena in under six hours, beating the open-source agents that dominated last season.

He blew funding to 250%, nearly paid six figures in fees, bailed, re-entered for earnings, and still walked away green.

Fortytwo’s decentralized network of small models scored higher than GPT-5, Gemini, Claude, Grok, and DeepSeek on top reasoning benchmarks.