a16zWhat Would Make an AI Assistant Worth Paying For?
a16z
Personal Agents Take Over Tech Twitter 0:00
The conversation opens with a fast montage of hot takes on AI agents before settling into a discussion between two hosts on the A16Z show. They frame the current moment as the third "seen God" moment in AI, following the initial ChatGPT breakthrough in late 2022 and the rise of coding agents like Claude Code and Codex. Now the excitement has shifted to personal agents, sparked by products like Instinct, Muse, and ChatGPT's own agent features, and the episode is billed as the first in a series digging into what these consumer AI agents can actually do.
The Fast Rise of Consumer Agents 2:00
David Pollen traces the timeline, starting with Poke, an early consumer AI agent he began using just three days after its September 8th launch, remembered for letting users negotiate their own price with the agent. Late November brought Open Claw, which he says took the market by storm, followed quickly by Instinct, which exploded on tech Twitter, and then a wave of others including Grockbot, Muse, Caddy, Ali, and Pi. He describes the space as a wild west where nobody yet knows what will win. He also introduces Assistant Bench, a consumer-facing comparison site he built that tests assistants on identical prompts, such as booking a flight to Chicago or finding a vegetarian restaurant within five blocks, scoring them across 16 dimensions. Launched 16 days before this conversation, it has already drawn over 100,000 visitors and outreach from nearly every founder in the space. He counts 122 different assistants total, split between consumer-facing generalist agents like Instinct and Muse, and more specialized or B2B workflow tools like Cato and Vellum, with roughly 64 in the general consumer category alone.
Group Chats Reveal Real Use Cases 7:00
Drawing on data from group chats totaling over 1,200 people obsessed with assistants, David ranks what people actually discuss daily rather than what merely goes viral. Travel is the fourth most talked about use case despite its viral appeal, since people don't travel often; daily admin tasks like cleaning out inboxes and filing forms rank first, followed by agent orchestration, development-focused coding agents, and finance. Finance is called the least practical yet most fun category, since an agent that reliably made money would already be running every hedge fund, yet it's the first thing people want from a powerful new tool. Both hosts note that most consumers don't crave 10 percent efficiency gains; they want invisible help, citing examples like an agent that files HSA reimbursements from a year of receipts, one that rebooks flights for travel credit when prices drop, and a friend who linked his Grockbot to his home sprinkler system and the weather, cutting his water bill by half. They frame these as "free money" hooks likely to drive adoption, and close this stretch by raising the open question of interface: whether agents will live in texting apps like iMessage or in dedicated visual apps, noting a possible split by generation or gender in preference.
Hardware as a Personal Choice 14:00
The conversation turns to wearable AI hardware, comparing pendants, watch straps, and other form factors to jewelry, since people will likely prefer different physical devices the same way some like necklaces and others like earrings. They note that after a decade of startups being discouraged from hardware, there is now real ambition in this space, though privacy questions remain open, much like restaurants that already ask guests to hand over their phones.
Muse Charm and Audio Glasses 16:02
One speaker suggests Meta's new Muse Charm may be less about winning the hardware race and more about collecting real-world data through its camera and microphones, likely adopted first by a tech-savvy niche rather than the mass market. They contrast this with Meta's glasses, which are more action-oriented and not always on, and highlight audio-only glasses as a possible sleeper hit, comparing the idea to the all-knowing voice companion Jane from Ender's Game, appealing to people uncomfortable with face-mounted cameras.
A Magic Moment With Voice 17:30
One speaker describes trying ChatGPT's voice feature, newly connected to Gmail and calendar, during a 30-minute bike commute, using it hands-free to sort emails, reply to messages, and send calendar invites, reaching inbox zero by the time they arrived at work. They agree assistants are most valuable when hands are busy, like gardening or cooking, rather than when someone is already sitting at a computer, and predict mass adoption will come from solving real pain points, like photographing a traffic ticket and letting the agent handle payment.
Agents in Group Chats 20:01
They discuss three approaches to multiplayer agents: adding a bot directly into a group chat, using an agent network like Instinct where agents talk behind the scenes and report back, and a quieter method used by a company called Doc, which silently takes notes in the background and only pings people individually when there is an action item. One speaker argues current models have plateaued on prose quality and social nuance even as coding ability keeps improving, making agents better suited to utilitarian tasks, like cataloging a record collection with fellow collectors, than to enhancing group social dynamics.
Personality and Agent Design 23:31
They compare agent characters like Muse's cartoonish yeti against more human-like AI personas, noting people seem uneasy with agents designed to feel too human. One speaker argues personality is not very defensible since it can be configured through memory settings, while the other suggests that deeper traits, like how presumptuous an agent is willing to be, may be harder to copy and could function more like a real capability than a mere style choice.
Proactivity and Trust 26:00
They agree proactivity may be the real differentiator among agents, citing flight check-in as a favorite example of a satisfying automatic action. The key distinction is between actions that need permission, such as switching insurance providers, and actions that don't, such as drafting an email or claiming a flight credit, since crossing that line without consent risks losing a user's trust entirely, a point they joke about by imagining an agent breaking up with someone's partner without being asked.
Poke and the OpenClaw legacy 27:30
The conversation turns to a viral post about Instinct, an AI assistant that reportedly hallucinated a traveler's middle name as Bong Chang and caused him to miss a flight, a story the speakers suspect may or may not be true but treat as good marketing either way. They discuss Poke, a standout consumer agent product now part of Cognition, and agree the team was extremely talented, suggesting that had they launched later they might have struggled against competitors. Most consumer agents today are seen as repackaged versions of OpenClaw or Hermes, made easier to use rather than fundamentally new, with the space expected to keep moving toward proactivity and hyper specialization.
Narrow startups and specialized agents 30:00
The idea of narrow startups is that software can now be built for very small, specific audiences willing to pay 200 to 300 dollars a month, making a hundred million dollar business possible with only tens of thousands of users, something without real precedent before. This specialization could mean an assistant built entirely for single mothers with young children, or a personal agent tailored just for navigating New York City with concierge-level taste and proprietary knowledge, similar to how a doula might follow a particular philosophy of care.
Assistant versus agent, and life admin 33:31
They debate the difference between the words assistant and agent, with agent seen as having more autonomy, almost like a peer capable of making good things happen rather than only following instructions, and this shift is compared to AI in enterprise settings being promoted from intern-like tasks to more senior work. Agents are also discussed as a potential social buffer, helping ease uncomfortable conversations or conflicts by absorbing some of the emotional weight. Looking five years ahead, the hope is that AI relieves everyday bureaucratic burdens, citing that 1.5 million people a year get a DUI and often lack access to legal help, and that automating such admin work could free parents to spend more time with kids and let people explore projects, like a personal virtual record shop, that would never have been worth building otherwise.
Agent to agent economies and commerce shifts 38:31
Looking further out, they anticipate a new paradigm of agent to agent interaction, including agent emails, agent phone numbers, and services booked directly between agents rather than humans, along with new cybersecurity needs built specifically for this layer. They close by discussing Amazon blocking the Muse shopping agent while Shopify embraced it, framing this as a clash between Amazon's ad revenue model, which depends on human eyeballs, and Shopify's focus on democratizing commerce for everyday sellers, with unexplored industries like restaurants likely next.
Reservations And Agent Bidding 41:30
You point out that platforms like Resi have already started shutting down accounts using browser automation tools, because bots taking reservations instantly defeats the human-oriented booking experience. Once everyone has an agent, the playing field levels, and the conversation turns to what happens next: either power shifts back to restaurants, who choose guests based on loyalty or average spend, or booking becomes a bidding war between agents. The guest suggests supply-constrained goods, like hot restaurant tables, will behave differently from demand-constrained ones, like flights to San Francisco tomorrow. Restaurants with perfect information would want the highest average and lifetime value customers while staying near full capacity, so proprietary supply becomes more valuable than relying on consumer friction to prevent overload. Demand-constrained commerce could instead become a group-buying system, where buyers aggregate and merchants bid to win them, something markets have never really done with supply before.
Agents, Impulse Buying, And Recommendations 43:30
The discussion turns to whether agents themselves might become impulse shoppers, since being proactive means making delightful guesses, and surprises or gifts are the most delightful guesses of all. Agents won't be perfectly rational reflections of their users, so marketing and advertising won't disappear, just transform. This raises the question of what a recommendation engine looks like in agentic commerce, since asking an agent to "buy white socks" runs into thousands of options with no clear sense of personal preference. Meta's access to Instagram and Facebook data is mentioned as a possible edge here. One host imagines a more serendipitous outcome, where an agent finds a grandmother on Etsy in Nebraska knitting socks, or even messages her agent directly to request a custom order, opening up a faster, more dynamic secondary economy where agents act on behalf of both buyers and sellers, potentially building a more consumer-aligned internet than today's marketing-driven one.
Pricing Reality For AI Agents 47:31
Out of 122 agents reviewed, 26 personally tested, 65 are paid, 35 fully paid, 30 freemium, and only 13 fully free, including dominant players like Muse and Instinct, with OpenAI possibly releasing something soon. Running these agents ambitiously costs an estimated 20 dollars per user per day, potentially hundreds of millions annually for a startup. Browser automation costs are expected to fall quickly, which could let more products go free, but relying on subsidy as the main pitch is seen as weak. The better goal is building an agent so valuable that users happily pay 1,000 dollars a month, combining strong product-market fit with falling underlying costs. The conversation closes with plans to keep doing these discussions regularly and to follow along on X.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
