Agents Take Action. Commerce Gets Smarter. Physical AI Scales.

What happened
AI startup Hark unveiled Handoff, a “browser agent” that can autonomously navigate sites (Target, Walmart, OpenTable, LinkedIn, etc.) to complete tasks like shopping or booking travel. Handoff uses a post-trained model designed to predict actions (clicks, typing) instead of just next words.

Why it matters
The startup claims Handoff is much faster and far cheaper to run than leading LLMs, saying it outperforms GPT-5.5 and Opus 4.8 on browser tasks. If true, this could turbocharge agentic AI by making real-world automation affordable and scalable.

What’s next
Hark is refining its model pipeline and plans to pre-train Handoff later this year. A waitlist is open, with a full launch targeted by the end of summer.

What happened
Shopify reported that AI-enhanced search has “become a complement to search,” not a replacement. In its Q2 earnings, the company said AI-driven traffic and orders tripled year-over-year, helping it beat revenue forecasts. Shopify CEO Harley Finkelstein noted that buyers using AI tools are landing directly on products, boosting conversions.

Why it matters
This suggests AI search is already improving real-world commerce metrics. By understanding buyer intent more deeply than keyword matches, AI tools are sending shoppers straight to relevant items. For merchants, that means more sales — and for the industry, it’s evidence that integrating AI into customer-facing tools can pay off immediately.

What’s next
Shopify is extending AI support by building connectors to multiple AI models (Claude, ChatGPT, etc.) and coding platforms. It expects AI to play a larger role in transactions (even at checkout) as merchants adopt these tools. Continued AI-driven innovation could further grow Shopify’s traffic and revenues.

What happened
Google AI stalwarts Jeff Dean, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals left Google to found Discovery Loop, a public-benefit AI startup. Jeff Dean will serve as CEO. The team has raised funding (with Alphabet and major VCs as backers) to apply AI to scientific research.

Why it matters
Discovery Loop aims to automate lab R&D, using “high-octane algorithms” to run thousands of experiments in parallel. This could drastically accelerate innovation in fields like biotech and engineering by cutting out slow human iterations. The founders’ pedigree (they built Google Brain and critical Gemini model research) signals a big bet on AI-driven science.

What’s next
Discovery Loop will deploy advanced AI systems for “recursive self-improvement” and experiment design. Backed by Google as an investor, the team will build tools that autonomously generate and test hypotheses. If successful, it could transform how research is done — effectively making AI itself a scientist in the loop.

What happened
WindBorne Systems raised a $37 million Series B (led by Khosla Ventures) to commercialize AI-powered weather forecasting. The startup flies high-altitude balloons around the globe to collect hard-to-get atmospheric data, feeding it into deep learning models that can run on modest hardware.

Why it matters
New AI techniques allow complex weather models to run on laptops rather than supercomputers. WindBorne’s “balloon data + AI” approach yields more accurate forecasts than satellite-only inputs. It’s already working with the U.S. Weather Service, Air Force and Navy, showing strong government demand for its data. The funding validates a growing trend: using AI and novel sensor networks to tackle big physical-world problems.

What’s next
WindBorne will expand beyond government contracts into the private sector – e.g. commodity traders using weather data to predict crop yields and energy prices. The new funds will also help replace costly satellite comms with mesh radio networks and build out more sensor types (like oceanic buoys). The goal is to turn better forecasts into profitable services for industry.

What happened
Google announced a major AI leadership shakeup. Demis Hassabis stepped down as DeepMind CEO to become Alphabet’s Chief Scientist and the Chairman of Google DeepMind, focusing on long-term AGI work. Koray Kavukcuoglu (DeepMind’s CTO) was promoted to SVP of Google DeepMind, taking over day-to-day operations (including the delayed Gemini model). Meanwhile, Google veterans Jeff Dean and Sanjay Ghemawat left to start an independent AI R&D nonprofit (Discovery Loop) – with Google and Alphabet investing in the new lab.

Why it matters
This split research from delivery within Google’s AI arm. Hassabis will pursue high-level science and AGI strategy, while Kavukcuoglu will “run the machine” on practical model development. It also signals urgency: Google gave those departing stars (Dean/Ghemawat) funding and support to continue AI innovation externally. The moves come as DeepMind’s flagship model launch has slipped, and show Google doubling down on both fast shipping and long-term breakthroughs.

What’s next
Google needs to deliver the promised Gemini 4 model to justify these changes. Developers and partners on Gemini should track version pins and deprecation notices carefully. Meanwhile, the Discovery Loop founders – now essentially in Google’s orbit as investors – will push on automated science. The industry will watch to see if this reorg helps Google stay competitive with OpenAI and Anthropic on advanced AI capabilities.

What happened
TechCruch reported Atoms (formerly CloudKitchens), Travis Kalanick’s newest robotics/automation venture, announced ex-Uber finance chief Gautam Gupta as its CFO. This follows a recent $1.7 billion funding round for Atoms. Gupta had previously been Uber’s CFO under Kalanick and helped launch the rideshare IPO.

Why it matters
Kalanick is literally “getting the band back together” to execute his “bits-to-atoms” vision. By bringing former Uber executives into Atoms, Kalanick aims to leverage that team’s experience in scaling a transportation platform. The hefty funding and Uber’s own $100M investment underline the scale: Atoms plans to apply robotics to major industries (mining, food delivery, transportation, etc.) at global scale. This hire signals Atoms is moving into rapid execution mode.

What’s next
Atoms will use its war chest to develop robots for Kalanick’s target sectors. We’ll be watching for details on its mining and logistics robot efforts. Uber’s continued backing suggests potential partnerships or deployments in ride-hailing or delivery. As Atoms ramps up R&D, it’s emerging as a unicorn contender in industrial robotics.

What happened
Tech distributor Avnet partnered with autonomous system developer Weston Robot to unveil an AI-powered inspection robot for factories and facilities. Combining Avnet’s edge AI expertise with Weston’s robotics, the unit can autonomously navigate large industrial environments (factories, ports, energy plants) and perform continuous visual inspections for anomalies (leaks, overheated equipment, etc.). The system uses AMD’s embedded AI chips to run neural nets on-device, enabling real-time analysis even offline.

Why it matters
This “Physical AI” platform shows how robotics and AI are converging to enhance safety and efficiency on the ground. Instead of manual spot-checks, a single autonomous robot can continuously scan for safety violations (PPE, intruders), equipment faults, or other risks. By processing data at the edge, it offers low-latency detection and can reduce costly downtime. As facilities grow in size and complexity, such systems promise a new level of proactive maintenance and incident prevention.

What’s next
The companies will pilot the inspection robot with select customers in manufacturing, logistics, utilities, etc. Beyond anomaly detection, expect the platform to be extended for tasks like inventory tracking or emergency response mapping. This collaboration signals a trend: AI moving from the screen into autonomous machines that act in the real world. If successful, more industrial operators may adopt robots like these to protect critical infrastructure.

AI is moving from generating answers to taking action across browsers, laboratories, commerce, and physical systems. The real advantage is shifting toward models that are cheaper, faster, and specialized enough to operate inside real workflows. The next winners will not just build more capable AI—they will turn intelligence into repeatable action.

Build a Shopping Agent Brief

Use ChatGPT to create instructions for a browser agent shopping for a specific product.

Include:

Then ask it to compare three products, explain the tradeoffs, and recommend one.

Insight: browser agents become useful when intent, permissions, and purchase boundaries are explicit.