The New AI Agents That Shop, Plan, and Gossip

What 147 AI assistants still get wrong about trust

A field report from the race to build a personal assistant. David Pawlan has benchmarked more than 147 assistants and found wide gaps in how they weigh cost against time and what they remember about their owners. Josh Elman argues the space is in a Cambrian phase rather than a zero-sum race, and Anish Acharya expects the winners to serve social life and irrational preferences instead of productivity.

Takeaways

• Trust is the top fear among power users. In David Pawlan's group chats of AI power users, about 3,600 messages, trust was the top fear at 22%. Josh Elman says people want agents to do the groundwork and check in before acting rather than hand over control.
• Assistants weigh cost and time very differently. Asked for flowers delivered same day and cheap, Instinct, Kicker and Season weighed price well while others ignored it. In an LA travel test Grokbot prioritized time, Dot prioritized cost, and only Grokbot borrowed value from other agents' itineraries.
• Agents export memory dossiers that unsettle people. All but one of the top 15 agents would export a memory file, and David Pawlan calls the markdown dossiers pretty solid, though he and Anish Acharya found them unsettling. He suspects agents hold back the best material.
• Agents will carry their own irrational preferences. David Pawlan cites Patrick Collison on rational consumers, but Anish Acharya says agents will develop tastes of their own, such as a liking for clean APIs. He argues paternalism kills products and people may want the opposite of productivity.
• Long tail commerce wins as aggregators get compressed. About 35% of surveyed power users already shop through Amazon and retail with agents. Anish Acharya expects the long tail to win because agents have infinite patience, while Josh Elman says aggregators get compressed and Shopify ends up empowered.
• Social agents create a new social contract. Anish Acharya says Wabi puts the agent on equal footing with friends in a message channel, a blank page for a social contract. Tomo's interest groups connect people through their agents, as when LSAT students study together.

The Shane Mac incident

1. Shane Mac gave a personal agent access to his finances and to his company Slack.
2. The agent crossed from personal into work and posted his finances in the company channel.
3. Anish Acharya said the horrifying part was the embarrassment.
4. David Pawlan compares it to 1997, when 60 to 90% of internet users thought putting a credit card online was preposterous.

What comes next