Andrew Winter

I feel like the awkward Catch22, and really my main gripe, with personal agents right now that's not really talked about is that there’s still not a great way to measure adoption/impact/success, or any other proxy for relative value vs. the competitive set and ‘worthiness of more capital investment’, besides obviously legacy metrics like product engagement and usage. And so Muse/Instinct/Grokbot/etc. are forced to care about and optimize for me opening the app and sending more messages, to the point where they are sending me constant notifications to get me to keep opening the app. Replacing my email inbox notifications with agent notifications is a neutral experience at best, and honestly it’s probably worse. Leaving a message unread is cognitively much more draining than leaving an email unread. But the promise of the personal agent is to make it so I never have to see notifications ever again. The software notification should be a relic of the past. I simultaneously want to pay $200/month (or much more) for a world-class agent product, and also literally never open the app or interact with the agent at all, basically ever. My agent should be like my tax guy, where once a year I sit down and spend a few hours with it tracking progress, signing documents, setting guardrails, and aligning on priorities for the next year. The current agent paradigm is basically just all of us subscribing to the same basic Mint plan to manage credit card spend. Where is the UHNW offering? Who's building the $10k/month agent that I never have to talk to? I want the full-self driving agent, even if that means there's actually a guy in a room taking the controls at times without me realizing it, as long as I don't have to think about it or open another app.

I think this might be called Wittgenstein's Ladder (but if it's not I am officially petitioning to coin it as Winter's Elevator @johncoogan @tbpn), and it's similar to Ford / Faster Horse, but my version of this is that you originally needed stairs (or a ladder) to get from the 1st Floor of a building to the 2nd Floor, at which point you could build an elevator between floors. Once you have the elevator though, the stairs become obsolete. We needed the stairs to get to the 2nd Floor so we could make the elevator, but the elevator is not just 'a slightly better version of stairs', or a model improvement that still relies on the same underlying technology. The stairs literally do not matter anymore once you have the elevator. This is how you should be viewing the software industry today. Most software and most business processes built over the past 50 years relied on a legacy set of stairs, but the AI agents just got off the elevator.

I think this is really spot on and worth reading. Also very similar to how we approach AI Adoption at Won. A few additional thoughts: - The legacy 'technical' vs. 'non-technical' gap was never as large as people made it seem. In retrospect, more companies should probably have been investing more in engineering upskilling before AI. @Austen was early to this. - With AI, the gap is less about technical proficiency and more about 'builders' vs. 'non-builders'. More technical folks are certainly more likely to be builders, but it's a not a direct correlation, and you likely have builders popping up in places you never expected. - AI as 'The Great Bundler'. We've seen this already across tools, eg just using Claude/ChatGPT (often w/ connectors) instead of opening individual SaaS apps, but I expect we will see this more across human capital too. If you're a builder, past some level of domain proficiency, why would you just be a builder focused on Design, or Finance, or CX? Job titles and scopes are going to evolve rapidly and horizontally. - Maybe <5% of the workforce dedicated to building (i.e. Engineers) was too low, but it was probably never going to be >10%. Automation tools (Zapier, Make, n8n, etc.) were on a path to bring that number closer to 15-20%, but AI has blown past that and is likely approaching 50%. However, 50% is still not 100%, hence we should still be realistic about what the KPIs for AI adoption might look like. - From an ROI perspective, AI feels much more reasonable than teaching everyone at the company how to code, but actually once you knew how to write code the cost was basically just the upfront investment. AI is different, and the investment cost (token spend) is recurring. If you're not careful, AI adoption can become a runaway train. - Just as an observation, defaults and herd behavior are extremely powerful. It feels like upskilling was generally default 'Off' for over a decade. It is now default 'On'. Defaults are also, as they are in this case, never precisely optimal, and they come with tradeoffs. - Plus a few diagrams I'm riffing on while thinking through this.