Key Takeaways

  • AI agents are pushing us into an "always-on" world, demanding unprecedented personal data and delegated authority.
  • USV General Partner Mike Mignano argues users will increasingly demand explicit alignment: agents must work for their specific incentives, not the underlying model's or lab's.
  • Harry Stebbings challenges this, suggesting user privacy concerns often give way to convenience, citing past tech adoption like Apple Pay and online dating.
  • The debate centers on whether agents acting as a "second self" represent a fundamentally new level of trust and loyalty compared to previous technology.

The Disagreement: Loyalty vs. Convenience

As AI agents become more sophisticated, taking on tasks from managing calendars to buying products, a core tension emerges: whose interests do they truly serve? Mike Mignano, a General Partner at USV, believes the sheer depth of personal delegation will force a reckoning on agent alignment. Harry Stebbings, host of 20VC, argues that history suggests convenience will win out, just as it always has.

Mignano frames the challenge directly, recalling an idea he published on X: “Who is your agent working for?” He argues that labs building these models have an inherent incentive to make their models "smarter, better, faster." This might clash directly with a user's goal of having an agent work exclusively for them.

He says, “I do think that we're entering a world in which context is increasingly valuable for the labs and also for the application layer. Right. And so I think we will see products and businesses increasingly try to push us further and further along the edge to an always on world.” For Mignano, this "always-on" context, combined with agents performing highly personal tasks like buying items or sending messages, means users will become "a little more self-conscious about the incentives of the model they're using."

Stebbings pushes back hard on this privacy concern. He believes the convenience factor will trump loyalty for most users. “I have to say I think people will get incredibly comfortable handing over the keys in the same way that we got incredibly comfortable putting our credit cards online using Apple Pay, finding our husband or wife online.” Stebbings predicts this level of comfort “will be de facto in 5 years,” suggesting that the shift will be so natural, the privacy debate will fade.

Mignano counters that this time is different. “I think we have never handed over so much of ourselves to a technology before than we're about to do with agents... it wasn't out buying stuff for us, right? It wasn't out sending very personal messages to family members and loved ones. It wasn't a second self, right? It was it wasn't us.” The distinction, for Mignano, is the agent's capacity to act as a truly autonomous, delegated extension of the user, blurring the lines of identity and agency.

Who's Right (and When They're Wrong)

Mike Mignano holds the stronger position, especially for founders building true AI agents that act as a "second self." While Harry Stebbings correctly points to humanity's historical trade-off of privacy for convenience, the nature of AI agents introduces a qualitatively different dynamic. Past technologies, like Apple Pay or dating apps, handled specific, isolated functions. They didn't operate as a fully delegated, always-on proxy for our entire digital lives. The risk isn't just data exposure; it's misaligned action on our behalf.

Stebbings' point has merit for lower-stakes applications. For an agent that merely optimizes your music playlist or provides quick factual answers, the bar for loyalty is lower, and convenience will likely dominate. People will readily cede some data for a better experience. However, when agents handle finances, personal communications, or critical decision-making, the user's demand for clear, explicit alignment with their goals becomes paramount. This isn't about simply handing over a credit card; it's about handing over the decision to use the credit card in novel situations, with the agent potentially holding its own incentives.

What to Do With This

If you're building an AI agent, shift your product's focus from mere utility to explicit loyalty. Don't just make it work; make it work for them in a way that feels unassailable. This week, audit your agent's core purpose. If it performs any action that could be perceived as having conflicting incentives (e.g., suggesting a product from a partner, or optimizing its own learning data), design transparent controls and opt-outs. Consider how you would build trust if your agent was literally a human assistant given access to your entire life. That level of transparency and demonstrable loyalty will be the key differentiator for high-value agents, not just privacy, but trust in action.