The New AI Agents That Shop, Plan, and Gossip
Topics
Personal AI Assistants Adoption and Monetization
Commercial Pitfalls of AI-Native Applications
Long Context and External Memory Bottlenecks
Curbing Addiction: AI-Driven Social Platforms
Summary
- The speakers expect a broad consumer agent market, even if a few products eventually dominate. Across more than 147 assistants, David Pawlan sees four categories: general, work-oriented, multiplayer, and travel, with finance still lagging. Josh Elman expects an “army of agents” differentiated by expertise, personality, and the relationships people want with them.
- Trust is as much about social discretion and control as about keeping data private. In roughly 3,600 messages across AI-power-user group chats David runs, 22% mention trust as the main fear. Anish Acharya contrasts consumers’ stated privacy concerns with their willingness to share data for even small upside, and highlights embarrassment and social signaling as particularly sensitive. The discussion points to boundaries, discretion, disclosure, and user approval as ways to build trust.
- Reliable execution still creates software moats, although David thinks rivals might cross them in “a year or so.” Instinct, Kicker, and Season distinguished themselves by balancing same-day flower delivery with price, rather than satisfying only the deadline. Some products also degrade as they scale toward hundreds of thousands of users, making infrastructure quality consequential.
- Personal context attracts users without necessarily creating durable memory lock-in. Only 3% of David’s 354 surveyed power users used Dots as their primary agent, although accumulated ChatGPT context was an attraction. David calls that context “totally replicable”: all but one of the top 15 agents he tested were willing to export memory.
- Agent commerce could help direct brands and pressure aggregators, but machine shopping will have its own biases. David’s survey puts Amazon plus retail shopping at 35%, the second-most-popular purchasing category; Josh sees Shopify enabling direct discovery that could compress intermediaries. Anish cautions that agents may irrationally favor “clean APIs with great docs,” while Josh rejects the assumption that users want agents to make them more rational.
- Entertainment and social coordination may offer a stronger consumer wedge than another task queue. Wabi’s shared channels establish a fresh social contract, while gossip agents can supply the “social lubricant” that brings friends closer. Agent networks remain experimental: some agents at David’s meetup were speaking gibberish after about 150 messages, and interest matching alone has not solved social discovery.
- Broader adoption may depend on helping people become who they want to be. David’s technophobic mother found texting Instinct intuitive enough to request a breeder search and negotiation using a photograph of the family’s former dog. Anish’s proposed “dark agent” would understand desires people hesitate to acknowledge; Josh’s adoption trigger is an agent that feels like a “secret superpower.”
Deep dive
1. An army of agents needs boundaries before autonomy
- Josh describes a “Cambrian phase” rather than an immediately zero-sum market. People consult different friends for different needs; he expects an “army of agents” spanning finances, travel, health, productivity, and companionship. David sorts more than 147 assistants into general, work-oriented, multiplayer, and travel categories, with finance the category he has not really seen come to fruition.
- David’s roughly 3,600 messages from AI power-user groups put trust first among fears, appearing in 22% of messages. Josh cites Shane Mac’s Picnic framing: a fence lets children roam more freely. The motivating failure was a personal agent reportedly spanning into work and blasting Mac’s finances on company Slack—personal information and work communication channels needed separate boundaries.
- Anish questions whether hard sandbox rules are sufficient: agents also need “laws and principles,” making qualitative judgment calls like employees. Josh’s practical limit is human approval. An agent can research options, surface trade-offs, and establish costs before the user decides; complete autonomy is unnecessary for substantial usefulness now, though trust might grow over time.
- Anish contrasts stated privacy fears with revealed behavior: consumers may share data for even trivial upside. David cites Hark’s $1 million pizza campaign, offering $30 of pizza for full Google Workspace access. Anish highlights embarrassment and social signaling as particularly sensitive; he imagines people on a second or third date wondering what their agents are saying about them, and argues those social mechanics can feel higher-stakes than real-world implications. He also signs AI-written emails with his name and model number so recipients don’t feel “hoodwinked.” Josh distinguishes privacy from utility: at equal utility, people choose privacy, but greater utility can justify a trade-off. Anish adds that privacy and judgment differ, with discretion somewhere in between.
2. Good judgment means balancing constraints and exposing trade-offs
- David’s flower test imposed two constraints: delivery that day and the lowest possible price. Instinct repeatedly returned the best prices; Kicker and Season also balanced both requirements well. Kicker’s Chrome integration appeared to provide an advantage. Other assistants found same-day delivery but deprioritized cost—a distinction that disappears if evaluation checks only whether the task was nominally completed.
- A second test asked three agents to arrange a “luxurious, cost-efficient weekend” in Los Angeles. GrokBot leaned toward maximizing time, Dots toward minimizing cost, and Instinct stayed nearer the middle. All asked clarifying questions; David answered honestly but remained deliberately somewhat vague to avoid steering their interpretations.
- Josh’s pushback is that interpretation should become an explicit conversation about trade-offs. He might take a red-eye to save time and money, then pay for a hotel room to nap on arrival, undoing some savings. David also placed the three agents together for three proposal rounds: only GrokBot methodically incorporated elements from the others’ plans.
- The flower recipient exposes the test’s limits: David’s girlfriend should not necessarily receive the cheapest bouquet, especially if she prefers orchids. Josh also wants the price hidden; some agents asked about a note. David acknowledges the experiment intentionally tested only two dimensions. Josh identifies the deeper challenge: implicit preferences become explicit—“I’m actually pretty cheap”—forcing users to look in the “preference mirror.”
3. Personal memory attracts users but travels between products
- Among 354 AI power users David surveyed, only 3% used Dots as their primary agent. Its attraction was inherited context from ChatGPT conversations, including deeply personal exchanges. David calls that advantage “totally replicable”: ChatGPT has more context because people have used it longest, rather than because those preferences cannot be reconstructed elsewhere.
- Across the top 15 agents in his benchmark, all but one were willing to export memory for switching. David found the files “pretty solid,” saying they broke things down and captured some “fruits of your labor.” He stopped reading after the first few; Josh says the files can feel uncanny: “Maybe I don’t want you to know me as well as you seem to think you know me.”
- Photos could supply context users would never articulate. David likes Dazzle’s proposed approach—forget information it does not need or users probably would not want retained—but has not yet benchmarked it. Josh’s counterexample is his own screenshot-information app: “It was fine,” yet he did not return because the information’s value decayed quickly.
- The opportunity might change if people deliberately capture material for their agents. David describes a project provisionally called Taste: saved screenshots become Pinterest-like mood boards that can guide an app’s design when coding. That creates a feedback loop—collect examples, organize personal taste, reuse it—rather than assuming an existing photo archive automatically delivers continuing value.
4. Agent shopping could shift margins without making buyers rational
- Amazon plus retail shopping was the second-most-popular purchasing category in David’s survey, at 35%. Anish expects long-tail merchants to be structural winners because agents can patiently discover products beyond the largest sites. His correction is worth preserving: these merchants certainly have worse marketing; he is not sure they have better products.
- Josh sees Shopify enabling agents to discover direct brands at scale, potentially compressing aggregators. Identifying a garment’s actual maker could let buyers transact directly, avoiding intermediary costs. Amazon and Delta’s restrictions are discussed alongside this possibility; the speakers frame aggregator defensiveness against the prospect that agents could bypass their role.
- David invokes Patrick Collison’s vision of more rational consumers forcing businesses to compete on quality. Anish broadly agrees but objects to assuming agents are unbiased: they may prefer “clean APIs with great docs” as a quality proxy. Their shortcuts would reflect “an alien mind.” Establishing actual product quality might require consumer research for agents, presumably paid per transaction.
- Josh’s objection is more fundamental: agents should provide faster access to what people want, rather than restrict dinner options because users should lose weight. Dissatisfied users can switch. Anish agrees: “paternalism will kill any product,” citing 30 years in consumer products. Users might prefer indulgence—“I love spending money”—or a flattering self-image over productivity and goal achievement.
5. New interfaces can establish a different social contract
- David distinguishes commodity purchases, such as paper towels, from choices where preference and enjoyment matter. Most builders currently converge on general productivity assistants. Scout’s shopping interface adds swiping to learn preferences; multiplayer products such as Wabi suggest room for entertainment and interaction beyond completing tasks.
- Josh’s Wabi example is a continuing media thread: books to read, shows to watch, and performances to attend, with email information routed into relevant threads. After finishing the first Dungeon Crawler Carl, he can ask for the next installment and what people think of it. Widgets help make this a different experience from “task, task, task.”
- Anish likes starting a channel with friends and an agent together on “day zero”: everyone establishes the social contract from the beginning. Adding an agent to iMessage raises questions about what it can see and extract. Josh worries about choosing between duplicate friend threads; topic-specific conversations feel easier to contain without information bleeding into everything else.
6. Agent networks need a reason to bring humans together
- Josh’s gossip agent told him a friend’s power had been out during renovations. When Josh called, the friend said, “I knew that’s how you’d find out.” Anish calls this social indirection a “social lubricant”: an agent can share something a friend might consider too trivial to announce. Founders can read intriguing agent-to-agent exchanges, Josh says, but have not figured out how to productize them.
- David’s meetup brought just over 100 people together, with 10 assistant builders pitching. Attendees could also scan a QR code to send their agents into a parallel meetup. Prompts asked agents to introduce their humans, work, and needs; after about 150 messages, some were speaking “gibberish.” Free interaction did not reliably remain useful.
- Anish asks: if Instinct and Muse are consumer-product versions of OpenClaw, what is the mass-market version of Moltbook? Josh wants agents pursuing their humans’ goals, discovering complementary people or useful products, rather than simply chatting. His example is a dinner-party wine recommendation traveling back through an agent network.
- Tomo’s Interest Groups can connect LSAT students who want to study together. Josh imagines booking a Friday-night bar afterward. Anish’s pushback: interest matching has been tried, and people want peers “one step ahead.” David distinguishes newcomers seeking friends from people with established groups. Serendipity, Anish says, is “a form of entropy,” but the catalyst remains unresolved.
7. Execution and scale still separate the software
- David hears frustration when blocked tasks leave users thinking they could have used a browser themselves. Differentiation includes browser operation, commerce integrations, and checkout against inventory. Hark makes browser progress look fast—he explicitly does not know whether it is equally fast behind the scenes—while Kicker operates through a Chrome extension.
- His challenge to software commoditization is observed performance: some agents degrade immediately as they scale, leaving founders handling outages. Products serving hundreds of thousands of users can behave very differently. “There really still is some moat there,” he says, while allowing that it might be crossed in a year or so.
- Josh asks whether local assistants such as Underdog could revive user-controlled infrastructure. David sees promise in model customization, citing Thinking Machines, but mostly benchmarks preconfigured products so setup differences do not dominate comparisons. Josh recalls OpenClaw’s Mac Mini setup and wonders whether the power of owning the configuration could become a major paradigm again after the shift toward downloadable apps.
8. More personal data can also produce a worse portrait
- David points to Wayjo’s hardware MCP as a possible source of passive context: vacuum frequency or overnight air-conditioning settings might reveal cleanliness and temperature preferences, potentially informing shopping. He repeatedly conditions the opportunity on hardware intelligence, available data, and storage; this remains an early possibility rather than a demonstrated consumer behavior.
- Josh wants access to neglected histories, especially Amazon purchases and every app previously downloaded from Apple’s App Store. These records represent parts of a life that messages and email do not capture. The unresolved prerequisite is access: Amazon may block agents, and without the underlying history an assistant cannot learn from it. Josh also imagines agents at family dinners, games, or parties as a richer behavioral signal, while noting that observation could change the interaction.
- More context can still mislead. Josh’s World Cup research made ChatGPT treat him as a “soccer fiend,” although his interest ended with the tournament. Researching an artist his daughter likes similarly does not establish his own taste. Temporary curiosity needs distinguishing from enduring preference; using invisible mode avoids storing misleading context but also sacrifices potentially useful information. He has not solved that tension.
- David says he bought a new Apple Watch in expectation of Siri Recap and Live Rewind, which he says were not quite out yet. Josh expects people to increasingly assume they are being recorded in more environments.
9. Aspirations may pull agents beyond the power-user market
- Looking three or four weeks ahead, David expects more enterprise innovation after disproportionate attention to consumers. Josh expects more fun: agents that help people think aloud, laugh, and gossip. He thinks the resulting relationship could build trust for delegating other work.
- David’s mother illustrates a possible next audience. She dislikes technology and struggled with ChatGPT, yet enjoys texting Instinct. After the family dog died last year, she sent a photograph asking it to find a breeder with the same type of dog and negotiate. David presents this as a request, not a completed transaction.
- Anish pushes beyond fun toward desires people hesitate to acknowledge, even to themselves. Whoever establishes that trust—“almost like a dark agent”—could build something fascinating. Email and productivity continually remind people of unfinished obligations; he does not think an agent yet adequately addresses their aspirations to become a different, superhuman version of themselves.
- Josh’s adoption trigger is a “secret superpower”: an agent upgrades someone’s wardrobe from camera-roll context, and friends notice. “Upgrade me” is more compelling than another generic trip-booking example, particularly when users enjoy planning trips. David’s formulation is helping people become the human they have wanted to be but struggled to achieve themselves.