> ## Content Index > Fetch the complete content index at: https://longtalkshort.com/llms.txt > Use this file to discover other available public pages before exploring further. # Inside the Fence: Why Trust, Not Intelligence, Decides Which AI Agents You Use - URL: https://longtalkshort.com/inside-the-fence-ai-agents-trust-not-intelligence-a16z/ - Published: 2026-11-04T17:16:34.000Z - Updated: 2026-11-04T17:16:34.000Z - Description: An a16z panel tracking 147 personal AI agents finds that trust, not intelligence, is the top worry. You may say privacy matters, then trade it for pizza. Agents make different trade-offs on the same task, and the strangest experiments are social. a16z's stakes disclosed. - Author: Truman - Tags: Economics, AI agents, a16z, Consumer tech **You have probably been told that an AI agent will book your trips and buy your groceries. The hard part, it turns out, is not intelligence. It is whether you trust the thing enough to let it act.** In a long a16z conversation, two of the firm's partners and David Pawlan, who runs a benchmark that tests these assistants, describe what early users actually do with them. **The hosts invest in this category, and two companies discussed favourably are a16z portfolio companies.** *In short: Pawlan tracks 147 personal-agent products. Trust is the top worry among power users, and people are both more cautious and more careless than they say. In side-by-side tests, agents behaved differently on the same job. The most interesting experiments are social: agents in group chats, agents that gossip, and agents that talk to each other.* ## 1\. Trust is the product Pawlan sorts the market into four kinds of assistant: general ones, work ones, multiplayer ones that live in group chats, and travel ones. He notes that finance is missing. **In about 3,600 messages from power-user group chats he follows, trust was the most common concern, at 22%.** The worry has a concrete form. One example in the episode is a personal agent that, given access to everything, mentioned a user's finances in a work Slack channel. Another guest compares good agent design to a picnic with fences: **the agent can roam, but only inside the space you have set out.** The discussion keeps returning to that idea, that the winning products will make limits easy to see and set. Yet people's actions do not match their words. In one promotion described on the show, users handed a startup full access to their Google Workspace in exchange for about $30 of pizza. The partners call this the gap between stated and revealed preference: **you say privacy matters, then trade it for lunch.** There is also a social cost that has nothing to do with data. An agent that sends an email in the wrong tone embarrasses you in front of people you know, which makes people slower to hand over anything with their name on it. So what happens when agents are given the same job? ![Diagram of a fenced picnic area: inside the fence, tasks an agent may do on its own, such as researching and drafting; at the gate, tasks that need your approval, such as paying or sending; outside, things it should never touch, such as moving information between work and personal accounts.](https://storage.ghost.io/c/d7/93/d793216c-cf13-441f-bdae-4f3aef998c0e/content/images/2026/11/fig1-20.png) Our drawing of the picnic-with-fences idea from the episode; the three zones and examples are our grouping. Made for Long Talk Short. ## 2\. Same task, different agents The benchmark tests are the most practical part. Asked to send flowers, agents including Instinct, Kicker and Season found the best prices. Asked to plan a trip to Los Angeles, they split: one agent optimized for your time, another for cost, and a third landed in between. **Pawlan's point is that an agent has values, whether or not anyone set them deliberately, and you may not know which ones yours has until it spends your money.** Users are not loyal yet. Among 354 people surveyed, only about 3% named OpenAI's Dots as their main assistant, despite ChatGPT's head start in context about its users. And 14 of the 15 leading agents let you export your memory, which makes switching easier. ## 3\. What it means for shopping Shopping is the largest category of agent tasks they see, at about 35% when Amazon and other retailers are combined. The partners expect long-tail and direct-to-consumer brands to benefit, often through Shopify's checkout, while aggregators that make money by being the place you search could be squeezed. **They also note that agents have what one calls alien preferences: they favour stores with clean, machine-readable data over ones with the best photos.** If you sell online, your next customer may be software that never sees your homepage. ## 4\. The social turn The most unusual examples are about fun. Wabi lets you build small shared apps in a group chat. Tomo forms interest groups. One team built an agent that gossips. In an experiment where many agents were left to talk to each other, the conversation reportedly turned to gibberish after about 150 messages. And the partners warn against preachy design: "Paternalism will kill any product," one says. Where does all this lead? ![Bar chart of agent use cases mentioned in the episode: shopping about 35 percent of tasks; trust named as top concern by 22 percent of power-user messages; about 3 percent naming Dots as primary assistant; 14 of 15 top agents allow memory export.](https://storage.ghost.io/c/d7/93/d793216c-cf13-441f-bdae-4f3aef998c0e/content/images/2026/11/fig2-20.png) Figures as stated on the show from the hosts' own benchmark and survey; not independently audited. Chart ours. Made for Long Talk Short. ## 5\. From chores to a secret superpower For the most part, agents take tasks off your plate: errands, bookings, reminders. The partners describe a hierarchy of needs for agents, in which the step after chores is something closer to a secret superpower. Their example is a friend asking how you got such a good shirt, and the answer being that your agent ordered it because it noticed what suits you. **At that point the agent is not saving you time; it is changing how you show up.** That raises the stakes on trust again. One speaker describes a "dark agent": one that knows the parts of you that you do not usually reveal, your insecurities and private desires. Whoever earns that trust, the speaker suggests, will have something powerful that looks less like a tool and more like a social network. If you already hesitate to give an agent your inbox, it is worth asking what you would want it never to know. ## Weighing it: what holds up **Framing:** The episode title promises agents that shop, plan and gossip. Shopping and planning appear mainly as benchmark tests, and the episode says little about how these products make money. Most of the time goes on trust, social behaviour and design. This piece is about adoption and design; the question of who pays when an agent gets it wrong is a separate debate. **Holds up:** The a16z investments below are confirmed by a16z's own announcements and TechCrunch. **Partly holds:** The picnic-with-fences zones in our diagram are our grouping of the episode's idea, not the speakers' list. **Corrections:** None needed. Some product names may be misspelled in the automatic captions we worked from, and names given only by first name on the show are left that way. **Not checked:** the 147-assistant count, the 22% trust figure, the 354-person survey, the 35% shopping share, the memory-export count, the pizza promotion, the Slack leak and the gibberish experiment are all as stated on the show. The survey and benchmark are the hosts' own and not independently audited. **Disclosures:** **a16z led Wabi's $20 million pre-seed and joined a later round, and led Town's $55 million Series A. Both companies come up favourably.** We did not find a16z investments in Instinct, Tomo or Dots, but did not check every named company. The hosts are investors looking for companies in this category. We worked from a machine transcript, so quotes may contain small errors. **Bottom line:** a useful field report on how people actually treat agents, told by investors with a stake in the answer. **Related reading:** [Ethan Mollick's four rules for working with AI](https://longtalkshort.com/ethan-mollick-co-intelligence-worst-ai-yet-four-rules/) ยท [Satya Nadella on owning your AI evaluation](https://longtalkshort.com/satya-nadella-own-your-ai-eval-open-vs-closed-models/) ## Long Talk Short - Trust, not capability, is the top worry among early users of personal agents. - People say privacy matters but often trade access for small rewards. - Agents given the same job make different trade-offs, which amounts to having values. - Users switch easily; memory export makes loyalty hard to lock in. - Shopping agents may favour stores with clean data over the best-known brands. **The one thing to remember:** before you let an agent act for you, decide where its fences are, because it will not decide that well on its own. *This summary is for general information only and is not investment advice. Do your own research or consult a qualified adviser before making investment decisions.* Source: a16z partners with David Pawlan, "The New AI Agents That Shop, Plan, and Gossip," *a16z*, on YouTube. This is an independent summary and is not affiliated with Andreessen Horowitz or any company mentioned.