The biggest challenge facing AI agent companies is that they're incredibly hard to adopt. Most people simply struggle to incorporate an agent as their primary work interface. In other words, there's a mismatch between the form these AI agents take and the context in which they're expected to be used.
For instance, I've had access to an AI software engineer named Devin for a few months now. The first time I used it, I had a four-hour session that felt almost magical. Watching Devin automate the role of a software engineer was seriously impressive. I threw a few tasks at it — like building small ideas I had and contributing to an existing codebase. It was excellent at creating new apps from scratch but stumbled a bit when it came to integrating into existing projects. That’s not entirely Devin’s fault though — it’s not a mind reader, and I probably should’ve provided more detailed prompts. What stood out to me during that session was the clear mindset shift I needed to get the most out of Devin. Within the Devin interface I wasn't an engineer, but a technical product manager. It felt like I was commanding a team of junior engineers to go and do my bidding. I distinctly remember feeling excited about this potential new workflow.
And then… nothing really changed. I thought about using it the next day but ended up skipping it to fix some urgent bugs in one of my projects. Sure, I could've used Devin to do it, but it would've required the extra effort of crafting the right prompt and monitoring the task. I already had all the context in my head, so it was quicker to just open Cursor and knock it out myself. This happened repeatedly over the next few days until I forgot about Devin entirely.
I had a similar experience with Minion, a general purpose AI agent product. Minion is an iOS app that allows you to command an agent to automate tasks that you would normally complete in a web browser, such as booking a restaurant or paying a parking ticket. The UX was engaging, and I enjoyed watching it attempt to book a restaurant for me. But I churned almost immediately, simply because I didn't have a pressing need for it at that moment. A few days later, when I needed to book a flight, Minion crossed my mind briefly as I started typing "flights.google.com" into my browser. But again, I defaulted to doing it manually because, one, I was used to it, and two, I had all the context in my head.
So, what's the takeaway here for creators of AI agents? If need to improve the fit between form and context. Agents must either: Integrate into existing tools we use daily, or Mimic the "analog" way we currently do things
Cursor is a great example — it's an AI-powered code editor, and they recently rolled out their own autonomous agent, Composer. I use it all the time because it's built right into my code editor. Crucially, Cursor isn't marketed as an "agent product"; it just has an agent as a feature. That distinction might be the key to unlocking broader adoption.