Services in a World of Agents

Friday is a time for philosophizing. This week, I had a thought: it would be great if everyday services had some kind of MCP or open APIs for agents. Then I started expanding it from the specific to the general. 🤔

We clearly assume that agents will become an increasingly significant part of ordinary people's lives. Almost everyone will have something like Siri on their phone and some decent alternatives like the desktop ChatGPT for getting all kinds of things done through external integrations. In that scenario, traditional apps may remain the primary interaction channel only for consuming content, while whenever we need to do something, we can turn to an agent instead.

There is clearly little point in using an agent to watch cat videos on TikTok. But ordering groceries from a photo of a half-empty fridge with a skill that takes into account how and where you usually order them is a great idea. Or booking a haircut. Or reserving a table. Calling a taxi. Or comparing prices for something you want across different stores and ordering it where it is cheapest. I am not even making much of this up—WWDC showed roughly the same thing. But I want it to work around us already, or at least to see things moving in that direction.

Right now, the only general-purpose way to do all of this is through elaborate workarounds like Computer Use in ChatGPT, which is not the most efficient approach. If only there were some concise protocol for integrating an agent with a specific service.

In the most normie-friendly scenario, an app next to the agent on the device could simply respond to requests like “give me products and prices” or “place the order.” In a geekier version, none of it would be so tightly bound to a particular platform or agent, so you could connect it to your Hermes claws. Of course, many solutions have already been reverse-engineered and recreated by someone, but we are civilized people. Official ones are few and far between.

Unfortunately, what is actually proliferating is superapps whose main goal is to make sure you see something else while completing your task. It feels as though we are closer to a future where every other app builds its own agents inside. They will be a thousand times clumsier and more awkward, but at least they will not let you escape their ecosystem.

I want to believe that the spread of agents will force services to reconsider, because giving users the ability to connect your product to an agent could become a serious competitive advantage. I would even pay for it. 💳