Every year I participate in an auction-draft fantasy football league. For those that aren’t familiar with auction style drafts, instead of each team picking a player in sequential order — teams bid on and buy players with credits. Each team has the same set budget and same number of roster positions to fill. This year I wanted to built a tool that helped me better find players and value them as the draft unfolded in real time, using position scarcity, ranking, and the purchase price of similar player to help me understand the true value of a player.
A couple days later I saw the release of Grok Bot and realized, instead of building a tool to help me value players and make smarter draft choices, I could just give instructions to an agent and have it manage my draft for me, surface choices with rational, while I just provide feedback. Will the agent make good picks? Who knows — but what I do know is an agent is better equipped to find, and analyze huge amounts of raw information than I could ever be. Yet it’s not the same still want to participate in my draft because it’s a fun yearly ritual to do with friends, and my ego still thinks that I would make better decisions than an agent. The Grok bot would likely be a better, more effortless tool than whatever I came up with, however it would do so at expense of experience. There’s a quality of using an artifact that was created by humans for humans that explains why purely generative software feels empty.
It’s entirely possible that the future that we’re hurdling toward is that every piece of software is collapsed into a chat app and all UI is generative. We tell it to do stuff and it does it. Software’s intended purpose, at it’s core, is to take problems related to information and data, and make them human sized and manageable; it conforms large scale problems and tasks to align to human limitations. Agents don’t have those same limitations. Any task that would require a human to use software, an agent can just… do, and do pretty well. This isn’t to say that an agent can replace a human worker, but it can replace the software-related executional work of information finding, connecting, shaping, calculating, and reporting.
Replacing software with agents doesn’t eliminate the need for all software, for software that remans it changes what that software means, who it’s for, and how it’s designed.
Software as we know it is mostly a means to an end. No one wants to use Excel, but they do want to solve accounting problems. And not all software shares the same posture; some software makes a task more efficient, like managing sales leads, while others such as youtube shifts how and where something like video is created and experienced. Then there’s also a fascinating space where these two postures blend, like retail, which has both experiential and executional bends. An agent fired from a chat app can buy you a shirt, but there’s a lot about clothes that’s so much more than the clothes, and our software needs to be expressive of that tangential, human information. In general, when we talk about how AI will change the form factor of our interactions with software, I think we have a tendency to greatly overestimate how much of our software use is actually execution oriented.
What makes software different from other tools is that feeling and experience is critical to the fabric of its use. When the physical form factor of a tool is undifferentiated, experience is how you give a tool not just shape but meaning.
What we’re finding (and models like Astra are confirming) is that agents are not just great at finding data and turning it into information, they can visualize and give shape to that information too. I think agent’s ability to visualize information is a forcing function that will prompt us to better understand what humans need from our software and interfaces, and evolve our practice beyond the display and manipulation of information toward building a canvas for the expression of a problem and the creation of meaning and relationship with the tool. It’s human nature to want to make our world better; to see something and want to improve on it. That is often an exploratory exercise rather than an executional one. It’s one thing to want to optimize and execute on something known like my fantasy football draft, it’s another to understand a problem from first principles —something you can’t necessarily point an agent at and give it parameters to solve, because you don’t quite know yet where to point. We need software to do more than compile data into information on a surface.
The shape and posture of software needs to shift from an experience-shaped tool that use design to bring humans closer to problem execution, to a tool-shaped experiences that bring us closer to information so that we can develop insights and express solutions.
The fantasy draft tool I originally wanted to build was never meant to make decisions for me. What I wanted was to see player values more clearly so that I could make deciiosns. It would enable me to spend more time with the problem and not feel rushed or timeboxed to make a decision.
That’s the distinction I think matters as agents take on more of the executional work. We like to talk about how much time humans will have to go spend time with friends and family, participate in hobbies and generally spend more time ‘offline’. And I’m sure that’s all true, but as it relates to work and our livelihoods, what’s most valuable is humans are relieved of the weight, pressure and planning associated with execution — we get back the freedom and flexibility and space to spend time with problems and develop understanding without feeling the burden of the how imposing on that processes.