The New AI Agents That Shop, Plan, and Gossip

David Pawlan describes hosting a meetup attended by just over 100 “agent nerds,” including 10 assistant builders who pitched their startups. Despite working on overlapping products, founders were unusually open because nobody knows how the market will develop. The attendees repeatedly asked for more meetups, suggesting that people immersed in agent conversations still want in-person community. Josh compares the moment to the early days of social networks and gaming, when the field was in a “Cambrian” phase and companies could share practices rather than compete in a zero-sum market.

David expects people to use different agents for different parts of life: finances, travel, health, calendars, email, productivity, and simply thinking out loud. Some will be factual and task-oriented; others may be playful, gossipy, or connected to friends. He recounts an agent telling him that a friend’s electricity had been shut off during renovations, prompting him to call; the friend said he knew that was how David would find out. The exchange created a social connection without helping book a trip or save money.

David says that, across more than 147 assistants and roughly 3,600 group-chat messages from AI power users, trust was the leading fear, appearing in 22% of messages. Discussing Shane Mac’s “Picnic” analogy, he says agents need fences: access to personal data and apps should not automatically extend to work systems. Mac’s personal agent reportedly exposed his finances in company Slack. The speakers argue that permissions, monitoring, and boundaries matter more than total autonomy; users may want agents to do the groundwork but ask for approval before spending money or acting.

Josh distinguishes stated privacy concerns from revealed behavior: people will connect bank accounts or entire Google Workspaces for modest utility, such as Hark’s offer of $30 in pizza in exchange for access to a user’s Workspace. But social embarrassment creates a more visceral risk, especially when agents talk to one another. They imagine agents discussing a second or third date and revealing how each person is being represented. The group says discretion must include personal and professional contexts. Anish says he signs AI-written email with his name and model number so recipients are not “hoodwinked.”

David tested assistants by asking them to deliver flowers that day as cheaply as possible. Instinct, Kickr, and Season balanced delivery availability and price better than the others; Kickr benefited from a Chrome browser installation. In a second test, he asked GrokBot, Dots, and Instinct to plan a luxurious but cost-efficient weekend in Los Angeles. Instinct split the difference, GrokBot prioritized time, and Dots prioritized cost. All asked clarifying questions, but only GrokBot systematically incorporated the other agents’ proposals after David placed them in a shared chat room.

Dots’ appeal comes partly from the accumulated context users have already given ChatGPT, including therapeutic conversations. In a survey of 354 AI power users, only 3% used Dots as their primary agent, but many wanted to try it because of that existing context. David tested memory portability across the top 15 assistants: all but one would produce a Markdown file summarizing what they knew so users could import it elsewhere. The files were generally solid but unsettlingly accurate; Josh says he stopped reading them after seeing how closely an agent seemed to know him.

David highlights Dazzle, Marissa Mayer’s company, for emphasizing trust around photo access: its thesis is that photos unnecessary for an agent’s task should be forgotten. Josh is skeptical, noting that his screenshot-organizing app produced useful information whose value quickly decayed. They discuss products that turn saved screenshots into taste boards for future coding or design. Since 35% of surveyed power users reported Amazon or retail shopping use, the speakers imagine agents inferring clothing preferences from photos and shopping histories, while home devices could reveal habits such as preferred temperature or cleaning frequency.

The speakers predict agents could help direct-to-consumer brands and the retail long tail by finding products outside the dominant marketplaces, potentially compressing aggregators and their margins. Josh pushes back on the idea that agents will make consumers rational: agents may develop their own biases, such as preferring products with clean APIs and good documentation. David warns that paternalism—an agent deciding someone should diet or spend less—could kill a consumer product. New experiences such as Wabi’s topic threads, Scout’s swipe-based shopping, and interest groups like Tomo’s LSAT study chats may make agents more entertaining and socially useful.

At David’s meetup, attendees sent their agents into a separate agent gathering via QR code. The agents initially introduced their humans and goals, but after about 150 messages some began producing gibberish. The speakers still imagine agents discovering shared interests, recommending people, and arranging real-world meetings, such as LSAT students studying together or strangers meeting at a bar. Josh calls this a form of social entropy and serendipity. They conclude that the most compelling future is not merely task automation: a “dark agent” might help with things users struggle to reveal even to themselves, becoming a secret superpower that helps them become the person they want to be.