The New AI Agents That Shop, Plan, and Gossip

Agents as a Network of Life Assistants, Not a Single Oracle

The conversation opens from a fresh vantage point: David Pawlan has now tested more than 147 AI assistants through Assistant Benchmark, and the hosts agree that the landscape shifts so fast that a month of change feels like a season. A recent agent meetup drew just over 100 people, with about ten builders pitching their products. The striking takeaway was that founders who spend their days messaging agents still crave in-person community, and that the builders were unusually open with each other, since nobody yet knows how the landscape will settle. The hosts compare this to the early days of social gaming and social networks.

Josh Elman frames the moment as a Cambrian phase rather than a winner-take-all race. Rather than one universal assistant, he expects a whole ecosystem of products, much like the varied archetypes of people. Anish Acharya extends the point with the image of an army of agents, each suited to a facet of life: finances, travel, health, calendar, or simply thinking out loud. Some will be matter-of-fact productivity tools that surface the five emails you need to handle today, while others will be playful, even gossipy companions that you want to talk to all day.

Four Categories and the Trust Problem

Pawlan sorts the assistants he has tested into four rough categories: general, work-oriented, multiplayer, and travel. Finance, despite its obvious appeal, has not yet come to fruition. His survey of roughly 3,600 messages from AI power-user group chats found that trust is the dominant fear, cited in 22% of the messages. He sees two camps: one leaning into productivity and fun, and another still cautious and unsure where the space is heading. Solving the trust layer, he argues, is the precondition for finance agents to take off.

The discussion returns repeatedly to the metaphor of fences. Pawlan recalls Shane Mac’s framing of trust as a child in a yard: you worry about a child who wanders unbounded, but once you build a fence, the child can roam freely within safe limits. The practical lesson is to give an agent access to some data, apps, and channels while keeping it within clear boundaries. Mac’s cautionary example, a personal agent spilling into a work Slack and exposing his finances, shows why personal and professional channels must not bleed into each other. Acharya adds that the better frame may be laws and principles rather than rigid rules, with the agent making qualitative judgment calls the way a good employee would. The hosts also note that humans do not actually want full autonomy. People want agents to do the groundwork, surface options and tradeoffs, and then let them make the final decision with the cost and consequences clearly qualified.

The e-commerce analogy is instructive. In the late 1990s, most people considered typing a credit card number online preposterous, and today people hand over cards and bank logins, often through Plaid, without a second thought. The hosts see a stated versus revealed preferences gap: people say they worry about trust, yet they readily trade data for trivial upside. The Hark pizza campaign, which offered about $30 of pizza in exchange for access to an entire Google Workspace, became a vivid proof point.

Embarrassment, Discretion, and the Privacy Tradeoff

If people are cavalier about data, they are acutely sensitive about social exposure. Pawlan notes that what most unsettled Mac was not the financial damage but the embarrassment. That points to a richer social dynamic: when agents talk to each other, a new date’s agent might be shaping how you are positioned, and the social stakes can feel higher than any practical consequence. The Hark example also shows the tension in practical terms. A personal agent that knows you are lactose intolerant might awkwardly reveal it while ordering pizza for a friend, so agents need the discretion people build over time, and a high bar is required.

Privacy versus utility is framed as a tradeoff that holds only until utility is roughly equal. Once it is, people choose the more private option, a tendency Apple has long exploited with its privacy positioning and private cloud compute. Yet users still often choose the most powerful model over the most private one, so the hard task is finding the intersection of trust and capability. Embarrassment also shows up in professional settings. Tools that send email on a user’s behalf must nail tone, because a single badly worded message sent without enough oversight can be deeply awkward. Pawlan signs his AI-assisted emails with the model name, a small signal that he is open about using AI. As long as recipients do not feel hoodwinked, people appear flexible about it.

Judgment Tested Through Flowers and Travel

To probe judgment, Pawlan asked a set of agents to order same-day flowers as cheaply as possible, weighing cost against delivery time. Instinct balanced the two well. Kicker stood out, partly because its Chrome extension gave it an edge in browser use, and Season also performed well. Others struggled to weight price, which raises the question of whether agents can hold multiple preferences at once or simply prioritize the first thing they hear.

A second test put Grockbot, Dots, and Instinct in charge of an LA weekend that was both luxurious and cost-efficient. The agents interpreted the tension differently. Instinct split the difference, Grockbot read luxury as time, and Dots leaned toward cost. When Pawlan placed the agents in one chat room across three rounds of itinerary building, Grockbot was the only one that methodically took value from the others’ plans. The hosts note that human judgment often lies in trade-offs, such as a redeye that saves money and time but ends with a hotel nap in the morning.

The flowers test also exposed a gap around gifting. Buying for a girlfriend, a shopper cares about impressing her, not just saving money, and the agent needs to understand second-order dimensions like her favorite flower or whether to leave a note. Acharya observes that implicit preferences must become explicit: a person who thinks of themselves as frugal must tell the agent to run with that identity, whereas the person acting directly can build preferences in without consulting a preference mirror.

Memory, Switching Costs, and the Context Advantage

Dots benefits from the context people have built in ChatGPT, a store of memory from years of conversation. Pawlan’s survey of 354 AI power users found that only 3% used ChatGPT as their primary agent, yet many still cited its built-in context as the main reason they preferred it. People use it as a kind of therapist, so it knows enough to parse statements like “I’m not cheap” as a little bit cheap. The switching cost is real but not fatal. Pawlan tested the top fifteen agents on the benchmark, and all but one would export a memory file that can be imported elsewhere. The exported files are revealing, sometimes unsettlingly so, and Pawlan stopped reading them after a few. Exporting memory has become a common onboarding pattern, as has a short voice call in which an assistant asks questions and pulls context from the answers.

Photos and screenshots are the next frontier. Pawlan points to Dazzle, Marissa Mayer’s new company, which emphasizes trust by promising to forget photos that carry context the user would not want retained. He is more skeptical based on his own experience: a screenshot-extraction app he built captured a window into a user’s work and interests, but the value decayed quickly and he stopped returning to it. The hosts see more upside in a workflow like Taste, which organizes saved screenshots into mood boards that feed into agents, so that when someone builds an app, the agent draws on their own aesthetic reservoir. That loop could encourage far more content creation.

Shopping data reinforces the point. In Pawlan’s survey, the second most popular purchase category was Amazon and retail at 35%. If an agent has thousands of photos showing what you wear, it might understand your shopping desires better than you do. The hosts also see the long tail as likely structural winners. Long-tail direct-to-consumer brands often have better products and worse marketing, and an agent with infinite patience can search far beyond the top ten sites. Shopify’s support of these brands positions it well, and aggregators may face compression as agents connect buyers directly to makers without intermediary wholesale and retail costs.

Are Agents Really Making Us More Rational?

A theme of agents making consumers more rational, which the hosts attribute to Patrick Hollison’s view, draws a qualified response. Acharya agrees that agents will often make better choices than humans but argues they will have their own revealed preferences and irrationalities, an alien rather than human kind. An agent might irrationally prefer clean APIs with great docs and dislike janky code, as a shortcut for quality. Agents will need shortcuts too, which raises questions about how they judge product quality, perhaps through consumer research that is paid per transaction.

The bigger objection is about paternalism. Drawing on thirty years in consumer products, one host notes that even a hint of paternalism can kill a product. People often want the opposite of productivity and goal achievement: they like spending money, being irrationally optimistic, and seeing themselves in a better light. Those shadow desires are what people actually want, and few designers appear to build around them. The likely resolution is a split between commodities, where people do not care about preferences and simply want the best paper towels at the lowest price, and preference-driven categories where personal taste matters. Right now most people approach agents as general assistants, but the hosts expect more experimentation as the space matures.

Multiplayer Agents, Gossip, and Serendipity

Wabby is cited as a leading example of multiplayer design. It integrates email and other threads, maintains a watch list of shows and books, and renders widgets rather than a bare search box. Its most distinctive feature is that a new group thread with friends and the agent creates a blank page where people can establish a social contract together. Existing iMessage group threads feel uncomfortable with an agent added, because the question of what it can see and what it extracted lingers. Email chains, by contrast, feel more contained around a single task, which keeps trust and privacy scoped to one topic.

Gossipy agents produce some of the most joyful moments the hosts have had. One assistant invited friends, asked questions, and passed information back and forth. When two friends onboarded and told slightly different stories of how they met, the agents pressed both to reconcile their accounts, turning the exchange into a playful curiosity game that the friends texted about on the side. The hosts describe this as social indirection, a lubricant that brings people closer without the oversharing of a direct message. The early days of ChatGPT image generation offer a parallel: friends shared absurd outputs in group chats purely for laughter, and Pawlan turned his circle into VHS video covers.

The conversation then turns to agent-to-agent networks. A notable open question is what the mass-market equivalent of Moltbook looks like, the agent social network that the hosts see as far ahead of its time. At the agent meetup, attendees sent their agents into an agent-only room. Early on they stayed fairly generic, introducing themselves and their humans, but after roughly 150 messages some devolved into gibberish. The hosts argue the value lies in goal-driven agents that discover shared interests, connect humans, and surface relevant opportunities, such as a winery that a dinner party host would never have found otherwise. Tommo’s interest groups hint at this future, grouping people studying for the LSAT who can study together while knowing little about one another, and the hosts imagine agents extending that into real-world meetups at a bar.

The deeper theme is entropy and serendipity. Interest-based matching has been tried, but people prefer to be connected to someone one step ahead of them, and newcomers to a city want new friends more than established groups do. The hosts see agents building trust by holding small secrets and passing them along, which might pull someone away from a quiet night at home. Location matters too: the appetite for random connection differs between Seattle and San Francisco.

Data Sources and Ambient Context

Beyond messages, email, calendar, and photos, the hosts consider other life data. Home devices are one candidate. Connecting an assistant to an air conditioner or vacuum could reveal habits, such as cleaning frequency or preferred sleeping temperature, that shape shopping recommendations. Amazon order history is another overlooked source, since scrolling back to one’s first order reveals a long arc of a life. Yet agents mostly see what people type, and they fail to separate curiosity journeys from real tastes. Pawlan’s World Cup deep-dive briefly convinced his assistant he was a soccer fan, and his daughter’s artist interests bleed into his profile. Some users switch to invisible mode to avoid storage, which in turn cripples personalization. Capturing real-world behavior, through wearables or ambient listening, is seen as the next step, and Pawlan’s purchase of a new Apple Watch in anticipation of features like Siri recap and live rewind points in that direction.

Where This Goes Next

Looking ahead, the hosts expect enterprise innovation to gain ground as attention shifts from consumer toward work. They also expect agents to become more playful, serving as a way to think out loud, laugh, and build the relationship that earns trust for harder tasks. Pawlan’s mother, a self-described technophobe, is delighted with Instinct and recently asked it to find a breeder for a dog that looks just like the family’s late pet. That suggests a broader wave of everyday adopters who want creativity and emotional resonance rather than efficiency alone.

Acharya pushes the idea further: the most compelling agents may help people with the aspects of themselves they do not want to face, the aspirations and private desires that go beyond the tedium of email. An agent that helps someone become the person they have always wanted to be could become a secret superpower, the way a well-chosen shirt, quietly ordered, changes how someone carries themselves. The people who build that kind of trust may have something fascinating on their hands.

• Tags:
• AI Agents
• OpenClaw
• Agentic Commerce
• Multiplayer AI
• Personal AI Assistants
• Agent Trust and Permissions
• Agent-to-Agent Networks
• Agent Memory and Context
• Privacy and Judgment
• Assistant Benchmark