8,000+ autonomous businesses have been born out of @polsia. Less than a year into @sapiom, we're already powering the 100,000+ agent runs per day that make that possible. For web search, @sapiom integrates with @Linkup_platform the web search API built for AI. Together, we’re rebuilding the infrastructure of the internet for agents.

Eleven months ago, we started building something because the infrastructure we needed did not exist. It's time for that to change. 🎱🖐️

Karp’s right that enterprises are “paying for tokens that create no value.” IMO, he’s partially wrong about the fix. His answer is discipline. Data sovereignty, more careful spend, tighter governance. But that puts the burden on the enterprise to manage something they can’t actually see. You can’t ask a team to exercise judgement over a system they have no visibility into. It's like sending your 4 year old off for the day with no guidance, no boundaries, no idea where they even went, and expecting them back with no scrapes, no bruises, and definitely no new pet snake in their pocket (swap to a cheaper babysitter and the kid may still come home with a snake). And watching from the front porch doesn’t fix it either. A camera pointed at the front door tells you the kid left. It doesn’t tell you where they went or why they came back with a snake. Watching a black box fail more slowly is still watching it fail. What actually changes the outcome is infrastructure that gives you real access at the task level: what ran, what it cost, whether it succeeded, how the routing decision got made. Not a dashboard. Control. That infra? It’s been built. And I’m damn proud of the @sapiom team for building it.

I left Shopify in August 2024 after 4+ years. $100B in GMV through Shop Pay. I left knowing exactly what world-class infrastructure looks like when it works. Before that, I built Earny, an automated agent doing price refunds for 4.5M users. It was acquired in '21. Two different companies. Same ceiling. At Earny, nothing like this existed, so we built it ourselves. At Shopify, the infrastructure was world-class, but it assumed a human was always the one pulling the trigger. Either way, the moment something needs to act on its own, the floor gives out.

When an agent fails in production, what's actually breaking? A recent VentureBeat survey asked enterprise teams running agents where only 17% said the model. What truly breaks is everything around it: state that doesn't survive a restart, steps that fail silently, no real record of what the agent actually did. That's the layer my team and I have built.