The New AI Agents That Shop, Plan, and Gossip

AI agents are still an early, non-zero-sum market, Josh Elman argues: people may use several for different needs rather than expect one to do everything. David Paulin groups the 147 assistants he has encountered into general, work-oriented, multiplayer, and travel agents, with potential roles in finance, health, calendars, and trip planning. Some are designed for playful or gossipy exchanges among friends, while finance assistants have yet to emerge strongly. The most compelling agents may combine useful tasks with conversation and entertainment, not just automate chores.

Trust is a central obstacle, particularly when agents handle money or represent users to other people. In about 3,600 messages from AI power-user group chats, 22% identified trust as their main fear. Limiting an agent to selected data, apps, and communication channels can establish “fences” that make it safer to grant wider freedom; keeping a personal agent out of work channels is one example. Agents can research options and tradeoffs but ask before consequential actions, since capability does not ensure sensitivity to human nuance. Rigid rules are not the only approach: models could instead be given principles and asked to exercise judgment.

Users may accept privacy costs for better service, but favor the more private choice when utility is equal. In one example, Hark offered a million dollars to people using its product to get free pizza; a $30 pizza required access to a Google Workspace account, including email and other data. The episode suggests that social embarrassment and how an agent speaks for someone can matter as much as practical outcomes. One speaker discloses AI-written emails by signing their name and the model number, reasoning that people are flexible if they do not feel misled. Trust could grow as it did for online credit-card payments, although a cited estimate says 60–90% of internet users in 1997 considered putting a card online preposterous.

Tests showed that agents differ in how they balance competing preferences. For same-day flowers at a low price, Instinct consistently found the best prices, with Kicker and Season also performing well; other tested agents found same-day options but struggled to prioritize cost. The test did not account for the recipient’s preferences, and raised the question of whether an agent can weigh several requirements or over-focus on the first one mentioned. For a trip to LA and a weekend that was both luxurious and cost-efficient, Instinct balanced time and cost, Grockbot leaned toward saving time, and Dots toward saving money. In a three-round group itinerary exercise, Grockbot was the only agent that methodically incorporated the other agents’ input.

Shopping is already a common use: 35% of surveyed users said they did some form of shopping in the “Amazon retail” category, meaning Amazon plus retail. Agents could match products to individual tastes and help smaller brands that have less marketing reach; Shopify is making products available to agents because it supports many direct-to-consumer brands. Whether agents will favor products with clean APIs, help consumers buy directly from makers, or simply speed up purchases remains an open question, as does whether they will make shoppers more rational. Scout uses swipe-style interactions to learn preferences, while commerce agents are getting better at checking inventory and completing purchases through browsers or direct integrations. But failed tasks can quickly send users back to doing the work themselves.

Personalization depends on retaining useful context from sources such as messages, email, calendars, photos, shopping history, and app downloads, without misreading it or preserving details that have become stale. A survey of 354 AI power users found that 3% used the discussed agent as their primary agent. Some tried Dots because of the conversation context accumulated in ChatGPT; a benchmark of the top 15 agents found that all but one offered a memory file for switching, often with detailed summaries users found unsettling. Photos and screenshots can reveal context people would not think to explain, but can be sensitive or quickly outdated. More data can also lead to false conclusions: World Cup searches led one assistant to infer a strong interest in soccer. Invisible mode protects curiosity-driven activity from being stored, but also denies the agent context that might help it understand the user.

Agents are also becoming social participants: some ask users and friends questions, pass information between them, and produce playful or gossipy exchanges that speakers described as a potential social lubricant. A shared channel can give friends and an agent a blank slate, while adding an agent to an existing iMessage group may feel intrusive; email chains can offer a more bounded, task-specific setting. Agents could introduce people with shared goals or turn interests into real-world plans, but users need to trust that the agent is acting in their interests and explain why a connection is worthwhile. Tomo, for example, has interest groups such as one for LSAT students, where agents can help members study together. At an event, agents sent to a separate meetup sometimes drifted into gibberish after about 150 messages, underscoring that useful social coordination remains a work in progress.