Let's call it the Battle of the Bots, Part 1: I decided to put two of the newest personal AI agents, OpenAI's Dots and Meta's Muse, through their paces, with the WNBA Finals providing a natural testing ground.

Why it matters: AI companies are racing to create the digital assistant we turn to for everything from shopping to travel planning. But there's still a gap between what these agents can do and what I'm comfortable letting them do.

• They also need to pass what we'll call "The Human Benchmark": that is, can they do something as fast, or as well, as a human? The answer is ... not quite yet, but the capabilities are still impressive.

The test: I asked each assistant to plan a fantasy trip to Atlanta to see my beloved Golden State Valkyries, organize a purple-themed brunch watch party and create souvenir programs.

• I also had them help me shop for Valkyries merch, which eventually turned into a side quest for WNBA basketball cards.

The cost: Muse is free for most uses, with optional $20 and $100 monthly subscriptions. Dots requires a ChatGPT Pro subscription, starting at $100 monthly, or access through a business or enterprise plan. More on Dots.

Where they shined: Both agents were surprisingly good at handling complicated, multistep assignments, such as looking up flights, finding a sign-up for a ticket presale and flagging a personal conflict.

Where they were creative, but imperfect: Both agents could turn broad instructions into surprisingly elaborate finished products, though neither was quite ready to replace a human editor.

Another creative test: I asked both agents to create a 1980s-style WNBA Finals souvenir program, then later to turn my photos from Game 1 into a keepsake.

• Dots' 38-page vintage program was more visually engaging, with player photos, graphics and extensive statistics. Muse's 35-page version was less attractive but had some fun touches, including a fictional computer ad and one for my imaginary brunch.
• Both programs contained questionable statistics. Muse later sent an unsolicited update correcting several factual errors.

Where they struggled: Actually doing things. I asked both agents to help me buy Valkyries merchandise, eventually settling on WNBA trading cards.

• Both helped compare options and prices. I chose Target, where my store credit card offered a 5% discount and free shipping.
• Muse attempted to complete the purchase but got stuck at Target's sign-in page. Dots instead directed me to the Target app, where my account and payment information were already saved.
• Dots' less automated approach turned out to be faster, even though both services support purchases with user approval.

The trust problem: Both agents could have done more if I'd been willing to give them greater access to my accounts and wallet. I'm not yet comfortable authorizing purchases of hundreds or thousands of dollars, or at a point of trusting them enough for consequential decisions.

The bottom line: For me, the sweet spot for these agents is doing the research and legwork while leaving me to make the final decisions.

• P.S.: I couldn't actually make the Atlanta trip or host the brunch because of other commitments. But I did end up buying three boxes of WNBA basketball cards. You know, for research.

A startup offered free pizza on Friday in exchange for letting its AI agent handle the ordering. After multiple attempts, I ended up with a $30 DoorDash credit, no pizza and an empty stomach.

Why it matters: AI agents promise to take everyday tasks off our plates, but they still have to clear a basic hurdle: making a task easier than doing it ourselves.

Driving the news: AI startup Hark advertised a $1 million giveaway, encouraging a download of Hark Pro and offering $30 in DoorDash credit as part of a promotion designed to showcase its agent's ability to order food.

• A number of customers reported experiences that were considerably less appetizing, including an investor who said it took 20 minutes, a user whose promotional credit didn't apply (who paid for the pizza out of pocket) and a startup founder who joked that she'd never worked so hard for anything.

Zoom in: I repeatedly tried to get Hark to order my pizza, without success.

• I did receive the promised $30 DoorDash credit, but the agent couldn't complete the order.

What they're saying: Hark acknowledged that the promotion strained its systems.

• "We're so happy the majority of people had a great night with a free pizza," a spokesperson said. "Obviously, we hate to have let anyone down and we've made updates on how we handle peak surge requests."

The big picture: Hark's experience illustrates a challenge facing the broader AI agent industry.

• AI agents can navigate websites, fill out forms and complete purchases, but doing so often requires authentication, permissions and human intervention.
• Those extra steps can make an otherwise simple transaction more complicated and, at times, easier for a user to complete on their own.

The bottom line: AI agents are getting better at doing things for us. But that doesn't necessarily mean they're making our lives easier.

Silicon Valley moves fast and breaks things, but the U.S. electrical grid moves slowly and requires permits.

• The explosive growth of AI is colliding with the glacial pace of building out new power infrastructure, prompting a new warning from Goldman Sachs that energy bottlenecks are set to stall the AI race.

Why it matters: Amid a wave of anxiety from Wall Street about the debt-fueled AI buildout, Goldman points to power as the biggest risk to trillions in spending in a new report shared exclusively with Axios.

Driving the news: Goldman expects global data center power demand to surge 170% between 2025 and 2030, with more than 60% of that growth coming from the U.S., according to the report.

• "Power is probably the forever bottleneck," Jason Tofsky, Goldman's global head of digital infrastructure investment banking, noted in the report.
• The bank estimates roughly $7.6 trillion in capital spending between 2026 and 2031 across computing, data centers and power infrastructure.

Zoom in: The physical infrastructure needed to support AI is moving far more slowly than the technology itself.

• The U.S. has roughly 2,600 gigawatts of proposed power generation and storage projects awaiting grid connections, more than twice the grid's existing installed capacity, though many projects may never materialize.
• The median time from an application for a grid connection to commercial operation is roughly five years.

Keep reading.

• Elon Musk says he will give one of his SpaceX shares to every "kid" in America next year. (Axios)
• Former Google CEO Eric Schmidt owns 1% of Anthropic, per Bloomberg.
• The AI trade is all the way back as Intel helped tech stocks rally, per WSJ.

Mady again. Will we start saying AI;DR, as in, AI, didn't read?

• That's the question on my mind after a viral Substack post points out that Amazon's famed six-page memo is no longer as useful because it's been turned into AI slop.
• One engineer began tagging AI-generated memos with "AI;DR" to indicate he wouldn't be reading them. His agent could just read and summarize those instead.