If AI in general and ChatGPT specifically were the buzzwords of 2023, then surely the buzzword of 2025 is "agents". Everywhere I turn, efforts to define it, build tools and infrastructure for them, build first approximations of them.
There are so many ideas of how to get to this future of many acting on the behalf of the few. Some like Perplexity, Dia and Yuturi believe they will come from the de facto tool to getting things done: browsers. I can see it: get great at navigating browsers to get things done the way that humans do, on the path to soon doing it on their behalf.
Others believe the future is via AI assisted virtual assistants - taking on whole tasks and projects like booking dentist appointments or ordering groceries with a back and forth. I can see that too - take a task and understand the parameters to do it well and learn.
These are all logical, reasonable approaches. It's just that, if there's one thing I've learned over the past decade it's: humans are anything but rational, predictable creatures.
And while I think it'll be much easier/faster to get to agents on professional domains, especially in well defined areas like coding, legal, medical records etc, I think personal agents are going to come down to understanding the nooks and crannies of how individuals do things.ย
There lies so many nuances - some people love the process, others need to start with a shortlist. Some want to feel challenged and others want it to just get done. And it differs task by task, category by category. What a user does with regards to meals might be different with booking travel. Or one partner in a household might want to do meals one way and the other partner another.
So how to even think about personal agents?
For me, it comes down to building something small but broad that can map a set of nooks and crannies... some fundamental space where we've not only built the effective "task path" for the agent to take, but the "trust path" that allows this interaction to ultimately end up in success.
Today, most of the focus is on the task path - what to-dos, what rational set of steps to execute. But without mapping the invisible counterpart, the trust path, there is going to continue to be a fundamental gap that prevents widespread adoption and ultimately, delight.