Anish Acharya, a general partner at Andreessen Horowitz, discussed the explosion of personal AI agents with David Pawlan, creator of Assistant Benchmark. Pawlan tested dozens of assistants on real tasks, covering email, booking, and financial administration. Both believe the most useful agents will become increasingly invisible, proactively checking in, seeking refunds, filing expenses, and handling daily minor tasks.

Pawlan's Assistant Bench compares assistants using the same prompts, such as booking a flight to Chicago or finding a vegan restaurant within five blocks, assessing success and follow-up on about 16 dimensions. He cataloged 122 assistants over a month, personally testing 26, and noted that many consumer products are merely rebranded generic agents. The discussion also covered how saving money tasks are superior to saving time tasks, Meta's Muse data and appeal, silent agents in group chats, the importance of proactivity as a moat, definitions of assistants and agents, and how Amazon intercepts Muse, Shopify opens up, along with an operational cost of about $20 per day for agents.

Acharya bets that the subsidy model is unstable and is willing to pay a high monthly fee for narrow, effective agents, which may be more robust than free generalists chasing millions of downloads. Interface candidates include apps, SMS, voice, and wearables; how much autonomy to grant and how agents place orders will shape the next layer of the internet. The episode was recorded on September 24, lasting about 51 minutes.

On the same day, the product showcase featured enterprise agents entering market shelves, car systems, and HR systems; this conversation shifted the focus back to personal administration. A daily cost of $20 implies that free generalists will rely on platform subsidies, while vertical expense reporting and flight price differences may encourage users to pay out of pocket.

In market mechanisms, what is purchased is an agent that can successfully complete real booking and reimbursement tasks, while what is sold is evaluation rankings and venture capital narratives. Funds enter the reasoning bill from consumer subscriptions or platform subsidies; beneficiaries are vertical agents that can connect to merchant interfaces, while those under pressure are shell agents that can only chat without changing flights. The event driver is the podcast launch, not a specific assistant winning.

The evaluation shifts from "like a person" to "was this ticket booked successfully." Being proactive without causing issues is closer to a moat than simply having another persona. Platform interception and openness determine whether an agent can place orders or only make recommendations.

Source: Public Information