How to Build an Agent-native Product | Mike Krieger

Models are amazing at piling on features and racing from blank page to a finished build, but they’re clumsy at pruning; the real craft now is knowing what to leave out and building the instincts that only come from usage.

Mike, welcome—cofounder of Instagram, now at Anthropic; what’s actually changed in product building from Instagram days to Anthropic, and what hasn’t?

Back then we spent a year on Burbn, then stripped it down to Instagram in a few months; today I can have Claude recreate Burbn in hours, filters and all, yet the hard part remains sharpening taste through real use, which is why true breakout consumer apps are still scarce.

It’s like growing a tree in a greenhouse—no wind, so it looks full but isn’t strong; rush the build and you skip the step‑by‑step intuition that makes great products, right?

Totally; YAGNI still applies, because overbuilding a version one creates a grid of features that’s hard to test and even harder to explain, like starting a show on the finale instead of learning the cast over episodes.

I did that with Proof—vibe‑coded a giant beast, then threw it out for a simple shareable markdown link and it took off; how are you reining in the temptation to add everything?

We rewrite early and often because today a redo takes days, not quarters, and we ship earlier to get contact with reality; Co‑Work was a ten‑day minimal cut that proved more than months of indoor tinkering would have.

I’m fixated on agent‑native design, where anything a user can do, the agent can too; Cloud Code taught me that pattern—how do you think about building for it?

Agents finally make computers feel like they just work by unlocking power normal people never touched, but every primitive needs to be editable by the agent, not just described back to you; we even package the agent‑native mindset as a skill so new prototypes start with that instinct.

Most models still think like old‑school engineers unless pushed; how do you teach them this new style?

Give them strong patterns and skills—like a current Cloud API skill—so they don’t argue from stale knowledge, and then build high‑fidelity harnesses because agent behavior will surprise you; robustness of the underlying primitives is the real job now.

Our bar is proof of use—I want a PR with a clip of you or your agent actually using it, not just green tests.

Same here: I ask Claude to prove it exercised the change and I expect proof of thoughtfulness, because the model makes lots of choices you didn’t, and we need to verify the primitives and their relationships hold up.

I had to onboard a SWAT team onto my vibe‑coded app and realized I don’t need to know every detail, but I do need enough to explain it—where’s that line now?

You can feel when a system is solid or a click away from collapse; think Instagram Direct v1 vibes versus a v2 that delivers messages reliably and still lets you push the system without the floor giving out.

How has hiring changed as models improved—we just brought on a lightly technical GM with strong product and writing instincts, but you can lose some low‑level rigor.

We still need senior systems thinkers for the primitives and to avoid prompt band‑aids that mask architecture issues, and we pair product teams with applied AI folks so we become customer zero for prompt and harness expertise.

Who owns UI and flow?

Designers who code are gold—some originate ideas, ship interfaces with polish, and pair co‑founder‑style with an engineer who paves the trail behind them.

So is it usually a designer or anyone with conviction paired with an engineer to smooth the edges?

What matters most is a conviction owner who will run through walls for the problem, flanked by rotating specialists; we review every two weeks to double down or release people back into the pool.

We keep teams tiny because one person can now hold more surface area, though there’s still a point where a feature becomes its own product.

Don’t scale early or you trade speed for coordination overhead; add a second brain when scope clearly warrants it or when fresh urgency will unstick the work.

AI moves so fast I toss half a product every few months; easier with one GM—how do you handle that churn, especially with enterprise?

Cultivate a culture of deletion and provide enterprise toggles while the core keeps moving, and migrate legacy features toward plugins or skills so they stop bloating the default surface.

For startups selling to enterprise, today’s modern stack goes stale quickly but customers want the old thing; what’s the play?

Set expectations that the train won’t stop, ship big rethinks with a transition window, then cut over—cycles are measured in months now, not years.

What’s your take on OpenClaw?

It spotlights the power and risks of tool‑enabled agents, and the real product question is finding the line between locked‑down and anything‑goes so it’s powerful, safe, and still feels yours.

Personal agents feel like mine because they’re named and tuned to me in a way the general model isn’t.

I like one named coordinator that delegates to sub‑agents so the run loop stays conversational, and the setup effort deepens attachment the way building your own kit does.

We’re even seeing a shadow org chart where each person’s agent inherits their specialty and earns trust from teammates.

That raises big privacy and disclosure questions, but the upside is an agent that applies what it’s learned from you rather than acting like a generic bot.

We’re out of time—where can people follow along?

I’m Mikey K on X; thanks for having me.

Smash like, hit subscribe, and buckle up for more AI and I—this ride is pure knowledge fuel.