Build Something Worth the Energy

A lot’s been posted about Y Combinator’s AI Startup School last week in San Francisco, but I want to cut through the hype and focus on what actually felt meaningful.

From the very start, it was clear that “the year of the agent” is here. Agentic AI is starting to take on tasks that feel more open-ended and fuzzy -- things like managing follow-ups across email and Slack without being explicitly told. If tech like Excel transformed how people worked with structured data, agents are now reshaping how we work with unstructured data and ambiguity - and the pace is insane.

Aaron Levie (Box founder) framed it well: historically, you can think about software trying to innovate the “nouns and verbs” each of us do every day. From 2008–2015, we solved a lot of the obvious ones -- write, send, search, share. But agents are expanding the verb set that software can tackle: now we can start to delegate, triage, monitor, refine. That’s a subtle but really powerful shift. Agents, unlike AI, are not things you slap on top of existing workflows -- they’re actually actors in the system. Karpathy made this weird but useful point: humans use GUIs, machines use APIs… but what do agents use? That question alone reframes how we design software. That’s why infra like MCP (shoutout Dedalus Labs) is gaining so much hype at the moment -- it’s like USB-C for connecting AI into products, a unifying layer that gives agents the scaffolding to act across apps, data, and environments with autonomy.

This shift in tech also seems to break the current software model. Sam Altman talked about a future where intelligence and energy are “on tap” -- local and always available. If models run on-device (à la Apple Intelligence), the need for APIs could shrink. And if just-in-time software becomes real -- apps generated as-needed by AI -- the idea of fixed, persistent products might start to fade.

As models move on-device and some apps become disposable, companies need new ways to stay in the loop. That’s why I suspect software-first companies are now thinking seriously about hardware. When software is ephemeral, defensibility thinking shifts -- it has to live in the parts that users keep coming back to: persistent products, intuitive interfaces, and systems that hold memory and context. The competitive edge moves from APIs to experiences -- the parts users come back to.

Perhaps unsurprisingly, this change is also shifting how teams are being structured. Andrew Ng mentioned a startup building their team out with 2 PMs per engineer -- more than flipping the 1:7 ratio I hear about. This isn’t because PMs suddenly became ten times better, but because engineers are no longer the bottleneck with AI. Execution is easier, and clarity is harder.

That gap -- the need for clarity and direction -- is exactly why design is starting to matter more in the future. Dylan Field mentioned how a lot of current design work is just “putting lipstick on a pig” -- cleaning up whatever AI spits out. But real design goes beyond that -- it’s becoming a discipline of craft and judgment. Knowing when to go deep, when to hold back, how to build systems that stand up when things get messy. That’s where value lives.

This is also reshaping what it means to be a SWE. Satya Nadella described the future SWE as more similar to a software architect -- someone who assembles and guides the orchestration of agents, balancing complexity, quality, and legal compliance (taking responsibility for it all). Roles are also blurring. The best engineers think like product people. The best PMs understand interaction design. Everyone needs to know when to keep things simple and how to make the complex possible. And just to be clear: this doesn’t mean coding is dead -- Andrew Ng was firm on this. If anything, being able to precisely express what you want from a machine is more important now, especially in a world flooded with autogenerated everything. Code is leverage, and it’s how you break from the default.

One last thing that stuck with me after talking with others after the event, and maybe it’s the most important: energy. Satya said, if we’re going to burn energy (literal or emotional), we should spend it on something that helps people. We’re living at a moment where machines are learning to mimic what it means to be human -- which makes it even more vital that we stay connected to what actually does make us human. Paying attention, caring, and building with others in mind is what that means to me.

If there’s one takeaway I’m holding onto, it’s this: Stay awake. Stay useful. And try to build something that’s worth the energy.