See pricing
The Personal Agent Race Is Here | Anish Acharya & David Pawlan
The a16z ShowAI

The Personal Agent Race Is Here | Anish Acharya & David Pawlan

The creator of Assistant Bench has tested dozens of personal agents. His verdict is that the winner will be the one you barely notice, and proactivity is the moat, not personality.

September 29, 2026 · 51 min listen · 13 min read · David Pawlan
0:00
–:––

Context

Personal AI agents went from experiment to one of the fastest-moving areas in consumer AI in a few weeks. Anish Acharya, a general partner at a16z, talks with David Pawlan, who built Assistant Bench, a consumer-facing site that runs the same real-world prompts (book a flight, find a vegetarian restaurant nearby) across many assistants and compares the results. For PMs, the episode is a live read on what consumers want from agents, where the interface and pricing questions are still open, and how commerce changes when software shops on our behalf.

The Big Idea

The winning personal agent will be the invisible one: it proactively handles tedious life admin without needing to be managed, and it never crosses the line into actions the user did not expect.

David's argument is that the general population does not care about being 10% more efficient. People hire an assistant so they do not have to manage anything, so proactivity, not personality, is where the defensibility sits.

Key Insights

Daily admin beats flashy demos

  • Virality vs usage: flight check-in gets shared the most, but in David's group chats (over 1,200 people across seven or eight chats) travel is only the fourth most discussed use case. People do not travel every day, so it works more as an attention grab.
  • Real ranking: number one is daily admin, such as cleaning out an inbox and filing forms. Number two is agent orchestration. Number three is development, such as using coding tools by text from a phone. Number four in the list is finance.
  • Caveat: these chats are all people in tech, so they are not representative of the wider population.

Cost savers may hook consumers

The hosts expect the breakout use cases to be cost savers, not productivity tools. Anish's point is that they will feel like free money, and Americans want to hear about free money more than saving time.

  • Going through a year of receipts to file HSA reimbursements.
  • Having an assistant contact an airline for a travel credit when a booked flight's price drops (David does this himself).
  • A friend who connected his Grok bot to his home sprinklers and the weather system, which David says cut his water bill by 50%.

Proactivity is the real moat

David sees massive defensibility around proactivity. The agent that acts on your behalf without being asked is the one that separates the winner from the pack. Anish adds that a lot of the magic comes from presumptuousness, and that if people stopped posting on X about what their agent did, it would suggest agents are not pushing hard enough.

  • Safe zone: proactive actions that need nothing from the user, such as drafting an email or getting a flight credit.
  • Ask first: actions that change something for the user, such as switching insurance providers because the current one costs too much.
  • The risk: David says crossing that line once means losing all trust with the user.

Personality is easy to copy

David does not think personality is a defensible mode. A user can tell the agent how to talk, and it saves that to its memory or soul file. He expects the leading general agent to be highly configurable.

Anish pushes back with a steel man. What looks like personality may really be capability. Some jobs need a highly presumptuous agent that makes mistakes 10% of the time as long as it gets things done. A finance agent should be highly cautious. How an agent breaks ties under ambiguity could be more defensible than its tone.

No single interface will win

  • iMessage: both hosts live there, and it feels personal because the user already lives in it. Anish says it holds a privileged place on his phone that a separate app does not.
  • Separate apps: David's example is Anish's wife, who prefers a visual interface for dream boards and travel.
  • Fragmentation: preference may split by generation (Gen Z toward texting, millennials and Gen X toward apps) and maybe by gender.
  • Other surfaces: an iPhone widget from a startup called Sky, and hardware such as the Muse charm.
  • Voice: an agent is most helpful when your hands are busy, such as biking, gardening or cooking.

Voice made inbox zero possible

David's second "magic moment" was ChatGPT voice. It had just added connectors for Gmail and calendar. On a 30-minute bike ride to work he categorized emails into labels, replied to emails and sent calendar invites, and arrived at inbox zero. Anish said it runs on a full duplex voice model and can see all your threads, so you can ask which three projects you started and what you are blocked on. Neither Muse nor Instinct had that yet.

Group chats are hard for agents

Anish finds that adding an agent to a group chat feels intrusive and can lower your social standing. David has seen three approaches:

  • Companion in the chat: the agent is added like another participant.
  • Agent network: as Instinct did it, agents talk to each other behind the scenes and update their users.
  • Silent listener: a product called Doc from XMTP, another a16z company, takes notes like Granola and pings people privately with action items.

Anish's view is that models have improved fast in verifiable areas like coding, but prose quality and bedside manner may have plateaued. So utilitarian agents (for example, cataloging a friend group's record collections) add more value than chatty ones. The design lesson from the silent listener is that the best group agent stays invisible.

Agents open commerce to new games

Both hosts think agent-driven purchasing breaks assumptions that current platforms rely on. See the trade-offs below for the Amazon, Shopify and restaurant examples. The recommendation layer is also unsolved: an agent told to buy white socks faces 20,000 options and does not know the user's taste.

Mental Models & Frameworks

Assistant versus agent

Anish's distinction is that an assistant can only do what you ask it to do. An agent has agency and can make good things happen in the world, like a peer in a social group. He compares it to enterprise AI, where agents started as interns and got promoted into higher-order work as their intelligence improved. The open question for consumers is what it takes for an assistant to be promoted in a person's life into more specialized and delicate work.

Narrow startups

  • Claim: software can now be very ambitious and very valuable for a small number of people.
  • Economics: people already pay 200 to 300 dollars a month for such tools, so a business could reach a very large run rate with tens of thousands of users.
  • Where the edge comes from: taste plus proprietary knowledge, such as an agent that follows one school of thought on childbirth or one that knows a city's best spots like a concierge.
  • Example: an assistant tuned for single mothers with children under three.

Supply-constrained versus demand-constrained commerce

Anish splits agentic commerce in two. Supply-constrained goods, like hot restaurant tables, may push power back to sellers who choose customers by high order value and lifetime value while staying near 100% capacity. Demand-constrained goods, like a flight to San Francisco tomorrow, may become a group-buying model where merchants bid to fulfill aggregated demand. He notes we have always aggregated demand but never supply in an auction that benefits the consumer.

Three webs inside the web

Anish sees the web as three parts: the commerce web, the application web and the content web. In his view the content web has been in trouble for a while. The commerce web has done well. The application web has been on a big upswing since coding agents arrived. The agent race puts new pressure on how traffic reaches commerce.

Trade-offs & Nuance

Consumer time versus consumer savings

The hosts cite Ben Thompson's point that most consumers want to spend time, not save it. Anish mostly agrees but says finance is an exception, since much of a US consumer's life is heavy administration, and many profit pools rely on consumer apathy. Agents could push those profits back to consumers.

Amazon and Shopify diverge on agents

  • Shopify: David describes it as focused on democratizing commerce, so agents buying fits its model.
  • Amazon: it makes its money on ad revenue. David says that without human eyeballs, the ad structure of commerce loses its point, so Amazon blocking Muse makes sense from its side.
  • Anish's view: Amazon is probably rightly betting it does not want to be disintermediated, since it would lose impulse shopping as well as ads.
  • Open point: agents are not perfectly rational, so Anish thinks marketing and advertising will not vanish but will have to be transformed.

Restaurant booking under agent traffic

David says Resy started shutting down accounts that used browser automation, since instant bot bookings defeat a booking platform built for humans. If everyone has an agent, the playing field levels. Restaurants may gain power to choose guests by loyalty or spend, or reservations may become bidding wars between agents. Anish adds that proprietary supply will matter more once consumer friction stops working as a natural brake on demand.

Free giants versus paid startups

  • The gap: of 122 assistants David tracks, 65 are paid (35 fully paid, 30 freemium) and only 13 are free. Muse and Instinct, the dominant ones, are both free.
  • Cost: Anish says an ambitious agent costs roughly $20 per user per day, which could reach hundreds of millions a year for a startup.
  • The bet: browser use should get much cheaper, as token prices have. A startup raising now would be betting on that within about six months, which Anish calls a game of chicken.
  • His preference: he does not like subsidy as the main value prop. He would rather see a $1,000-a-month agent that people want so much they pay for it.

Connector or general agent

Travel and finance are infrequent, high-value behaviors, so Anish wonders whether they become billion-dollar vertical connectors instead of general agents. David expects perhaps one independent travel winner, but mostly connectors into a general agent.

Common Mistakes

Mistake: Humanizing the agent too much

Anish notes that group chats feel uncomfortable when an agent takes the mic. David adds that AI-generated human characters make people uneasy, while a cute Yeti (Muse's character) has been embraced. The risk is that an agent trying to feel like a person feels weird in a personal space. The alternative is a less human design that stays out of the way.

Mistake: Crossing the permission line once

David says an agent that takes a direct action affecting the user without asking, such as switching insurance, can lose trust immediately. Anish jokes that the extreme case is an agent messaging someone, "I noticed you weren't that into her, so I went ahead and broke up with her." The better approach is to be proactive freely on save-time and save-money tasks, and ask first on decisions.

Practical Application

Score your product's proactivity

List every action your product takes for the user and mark each as "no user action needed" (like drafting or refunds) or "changes something for the user" (like switching a plan). Allow the first group to run automatically. Require permission for the second group, since David says one crossing costs all trust.

Start with cost-saving workflows

  • Do: pick one workflow where the agent can find money for the user, such as reimbursements or price-drop credits.
  • Why it works: Anish expects this to feel like free money, which is an easier hook than efficiency.
  • Check: ask whether the workflow is frequent enough. Travel is high value but infrequent.

Design for hands-busy moments

Test your agent flow in a situation where the user cannot look at a screen, such as commuting or cooking. David's bike-ride inbox-zero session shows voice can complete a chain of tasks (labels, replies, invites) in one sitting.

Make group agents silent by default

If your product adds an agent to shared spaces, prototype a listener mode like Doc's. It takes notes and pings individuals privately about action items instead of speaking in the group.

Model the unit cost early

Anish's estimate of about $20 per user per day for an ambitious agent means pricing needs a plan. Write down what happens to your margins if browser-use costs fall over the next six months, and what happens if they do not.

Questions to Consider

  • Which actions in your product change something for the user without a clear request, and which only save them time? Where does the second group sit relative to the trust you have built?
  • If a user could tell your agent "respond in this style" and have it remember, what part of your product would still be hard for a competitor to copy?
  • If shoppers used an agent to buy from your platform, which part of your revenue depends on a human looking at a page, such as ads or impulse purchases?
  • Would your users pay $1,000 a month for one agent, and what would it have to do so consistently that they never thought about it?

Bottom Line

Personal agents will win by being invisible and proactive, starting with the tedious admin and money-saving tasks people hate. The open fights are about interface, unit cost, trust boundaries, and how commerce works when software shops on our behalf.

Resources Mentioned

ResourceTypeWhy it was mentioned
Assistant BenchWebsiteDavid's consumer-facing benchmark comparing personal AI assistants on real tasks across 16 dimensions. It passed 100,000 visitors in its first 16 days.
Ben Thompson on TBPNInterviewCited for the point that most consumers want to spend time, not save it.

Tools & Products

Tool / ProductWhat it doesWhy it was mentioned
PokeEarly consumer agent with a sassy personalityDavid's first agent. He used it three days after launch and was struck by its negotiation-on-price gimmick.
InstinctConsumer agent that lives in iMessageBlew up on tech Twitter and uses an agent-to-agent network.
MuseMeta-linked consumer agent with a Yeti characterFree and dominant, with a newly announced charm wearable. Shopify embraced it and Amazon blocked it.
ChatGPT voiceFull duplex voice mode with connectorsLet David clear his inbox by voice on a bike ride.
Doc (XMTP)Silent group-chat listener that takes notesExample of an agent that stays invisible and pings users privately.
OpenClawOpen-source agent frameworkDavid sees most consumer agents as preconfigured copies of it.

People to Follow

David Pawlan

Creator of Assistant Bench. He has tested 26 of the 122 assistants he tracks and runs group chats of more than 1,200 people obsessed with agents. Follow him on X at x.com/DavidPawlan.

Anish Acharya

General Partner at a16z who hosts the conversation and pushes on defensibility, pricing and agent-to-agent commerce. Follow him on X at x.com/illscience.

Notable Quotes

"The general population does not care about being 10% more efficient." (David Pawlan)

"There is massive defensibility around proactivity." (David Pawlan)

"If you cross that line once, you lose all trust with your user." (David Pawlan)

"It's not about cost reduction. It's about possibility expansion." (Anish Acharya)

AI PM course

Everyone hears the same episodes.
Few can do what they describe.

Start for free