The true moat in personal agents isn't personality—it's daring to act without asking
Personality is configurable; what's genuinely hard to replicate is initiative—how far an agent dares to act without permission—and crossing that line breaks trust instantly.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
From zero to 122 personal agents in a single month
Poke launched September 8; David started using it three days later on the strength of its price-negotiation gimmick. By late November, OpenClaw exploded across the network, followed by Instinct, GrokBot, and Muse in succession. By the time of recording, AssistantBench had catalogued 122 personal agents—64 general-purpose, the rest split across verticals like travel and email. Industry consensus: nobody knows where this ends. Everyone is improvising.
— DavidThe finance use case is least practical but most enticing
In discussion groups of over 1,200 people, most-debated use cases rank in order: routine admin work (clearing inboxes, filing expenses), agent-to-agent coordination, development workflows (telling an agent to write or redesign code), then finance last. Finance is seen as least credible—nobody actually gets rich on a phone agent; hedge funds would have done this already—but also most irresistible, because people's gut instinct when handed an agent is ‘can you make me a million dollars?’
— DavidSaving money moves ordinary users more than saving time
Typical consumers ignore 10% efficiency gains. What truly moves them is silent cost arbitrage: organizing receipts from the past year into HSA reimbursement claims, automatically requesting travel vouchers from airlines when fares drop, connecting GrokBot to a home irrigation system tied to weather data—one friend's water bill fell 50%. These cost-saving moves that need no approval are seen as agents' easiest trust wedge.
— DavidMuse's pendant strategy might not be about winning hardware
David offered a theory circulating among insiders: Meta's Muse pendant is less a hardware play than a way to embed a 24/7 device with camera and multi-mic array that feeds real-world data to Zuckerberg's metaverse ambitions. He expects the device to stay confined to tech enthusiasts but to become a steady source of data on how the physical world actually works.
— DavidThe best agent in a group chat is the one that disappears
Dropping an agent into a group chat often feels intrusive and drains goodwill. David argues the most effective model is XMTP's Doc-style granola approach: the agent listens silently, jots notes, sends action items as direct messages to whoever's accountable, otherwise vanishes from the group entirely, and surfaces findings in its own document only when needed. Less friction, easier to embrace.
— DavidPersonality is configurable; the willingness to act without asking is the real moat
Poke became famous for acerbic wit, but both guests agree personality is a setting—one line and users reprogram the agent's tone; not a moat. The genuine barrier is audacity: how much the agent dares to decide for you without asking. They agree on the boundary: cost- or time-saving moves that need no approval can be proactive. Anything requiring a user to change behavior needs consent first, or you've breached trust irreversibly.
Amazon locked Muse out; Shopify threw the door open
Shopify integrated Muse this week. Amazon shut it down. David sees two business logics colliding: Shopify wins by lowering the friction for ordinary people to start selling. Amazon's revenue is advertising—the moment agents skip human eyes and buy direct, that ad engine fails. Anish adds a parallel: restaurant bookings face the same rift. Resi has started deplatforming accounts that use browser automation to snatch seats. If everyone uses agents to grab reservations, platforms may hand power back to restaurants to filter by spend, or spark agent-versus-agent auctions.
$20 per user per day is drowning out the rest
Of 122 agents David tracked, he personally tested 26. Among the 122: 65 charged money (35 purely paid, 30 freemium), only 13 entirely free—while Muse and Instinct, the current leaders, are free. Anish offered the team's cost estimate: doing this right costs roughly $20 per user per day; running it could burn billions annually. His thesis: instead of subsidizing free eternally, build something users want to pay $1,000 a month for, and win on genuine demand, not cost curves.
In their own words · checked verbatim
I think the general population does not care about being 10% more efficient.
David10:25
And so now his sprinklers are dependent on the weather and it's dropped his water bill by 50%.
David11:25
I have a hot take thesis that this Muse charm is actually less about trying to win the hardware game. And it's more about data collection in the real world for the meta team in terms of like how the real world actually functions, given it has a camera and multiple microphones.
David16:31
I was biking to work. Hands are busy. I want to get stuff done. And I just start talking to ChatGPT Voice. … And throughout my 30-minute bike ride to work, categorized my emails, submitted them to different labels, replied to different emails, sent out calendar invites, got to the desk, inbox zero.
David18:32
I think there is massive defensibility around proactivity.
David25:41
We were joking internally, like we're days away from an agent messaging someone saying, I noticed you weren't that into her. So I went ahead and broke up with her for you.
David27:47
You have Shopify that's hyper-focused on more democratizing commerce. You have Amazon that makes all of their money on ad revenue.
David41:15
Man, it's going to be really challenging. By our estimate, it's costing something like $20 per user per day to do this in a really ambitious way. Potentially hundreds of millions a year for a startup.
Anish48:32
Figures
| Total personal agents tracked | 122 (64 general-purpose) | 6:14 |
| Agent discussion group participants | 1,200+ | 7:18 |
| Water bill reduction from irrigation connected to weather data | 50% | 11:25 |
| US annual arrests for drunk driving | 1.5 million | 36:00 |
| Agents David personally tested | 26 out of 122 | 47:32 |
| Paid agents among 122 tracked | 65 charged (35 pure paid, 30 freemium), 13 free | 47:32 |
| Estimated operating cost per agent | $20 per user per day | 48:32 |
Glossary
- Narrow startups
- Startups serving a small, affluent audience with extreme specialization in their vertical.
- Presumptuousness
- An agent's inclination to make decisions and take action on behalf of a user without explicit permission.
How to listen
Entrepreneurs, investors, and anyone evaluating whether to build or back a personal agent.
32:00–35:00, the debate over agent versus assistant terminology and whether agents can improve social relationships.