The Smaller Shape of AI-First Companies

I’ve been thinking about what Anjan Katta said in his recent interview about AI companies - that the point isn’t building foundation models, but using AI to solve what makes companies break down. That most organizations fall apart, producing bad work, because they scale past the Dunbar number and accumulate bureaucracy, meetings, and overhead that suffocates actual work.

Traditional companies expand by hiring more people. Need a new marketing channel? Hire a marketing team. Want to test a new product line? Build out a whole division. Each expansion means more coordination overhead, more meetings, more cruft.

AI-first companies will expand and contract much more fluidly. Instead of permanent organizational arms, they’ll spin up AI agent teams for specific initiatives. Want to test that new marketing channel? Deploy a suite of specialized AI agents with deep expertise in that area. They can run campaigns, analyze performance, iterate on messaging - all coordinated by a small human team that sets direction and makes key decisions.

The company can suddenly have deep expertise in areas that would traditionally require months of hiring and onboarding. When that experiment is over, the “team” dissolves or pauses seamlessly.

This is already how I’m doing my own software engineering today: I spin up ephemeral copies of even large projects and set out on an idea with a few agents collaborating. I can iterate on an idea in 20 minutes and quickly get to a demonstrable success or dead end, which I then package up for others to review or choose to discard. I’m thinking while I prototype, starting out on paths with a kernel of an idea.

Branches and prototypes, growing and blooming or breaking.

Successful companies will do the same - using AI to enhance prototypes, agency, and curiosity. You’ll still hire dedicated teams when you decide to really invest in something. But the exploration phase - the risky, uncertain part where most initiatives fail - becomes much cheaper and faster. The core team can focus on what they’re actually passionate about and uniquely good at.

For this to work at scale, asynchronous collaboration becomes critical.

The whole team needs to be excellent at:

• Clear direction-setting - operators (both human and AI agents) need unambiguous goals and constraints to move forward in parallel
• Asynchronous feedback loops - Quick ways to review, redirect, and approve work happening in many parallel threads
• Documentation and knowledge sharing - Context that agents and humans can both access and build on (previously, previously, previously, previously, previously)

This isn’t just about time zones anymore. It’s about creating workflows where humans and AI can hand work back and forth efficiently, playing in the same space.

That said, I’m optimistic we’re moving toward something more collaborative than today’s fire-and-forget AI interactions. Instead of purely asynchronous handoffs, I envision AI agents that can participate more directly in our normal conversation loops - incorporating real-time feedback, asking clarifying questions, and behaving more like active collaborators. I imagine the conversation as work, where the back-and-forth dialogue between humans and AI agents will match the collaboration I have with real human conversations.

This significantly changes what’s possible for small teams. A 10-person company now has the ability to test product ideas across multiple markets, run sophisticated marketing campaigns, and analyze complex datasets - all simultaneously and with minimal investment.

The constraint shifts from “Do we have the capacity?” to “Do we have the clarity (of vision and the systems) to coordinate all this activity?”

The teams that master this approach will have the advantage. They’ll explore more opportunities, move faster on good ideas, and avoid the organizational bloat that makes companies slower and more brittle as they scale.

Strong companies in the future will be smarter and smaller. They’ll compete on speed, focus, and curiosity rather than raw headcount and bureaucratic process.

And I’m starting now: embracing the creative exploration that I can do with a team of agents. I already see my most productive weeks are those where I switch from raw implementation to modes of diagnosis and prototyping: fast-moving and focus-building.