Build for Agents, Not Humans: Why API-First Software Is the Only Path Forward
Aaron Levie, CEO of Box, just published what might be the most important essay in enterprise software this year. His thesis is brutally simple: AI agents, not humans, will become the primary users of all software in the future. And if your software isn't built for them, you're already dead.
From "Make Something People Want" to "Make Something Agents Want"
Paul Graham's famous Y Combinator motto — "Make something people want" — defined a generation of software companies. Build tools that are simple, easy to adopt, solve clear problems. That advice created billions in value.
Levie argues we need a new version: Make something agents want. In a world where every employee has dozens or hundreds of AI agents working on their behalf, those agents — not humans — will decide which tools get adopted. "Agents won't be going to your webinar or seeing your ad," Levie writes. "They're just going to use the best tool for the job, and you'll want it to be yours."
The API-First Imperative
The biggest practical implication is stark: if you don't have an API for a feature, it might as well not exist. If it can't be exposed through a CLI or MCP server, you're at a disadvantage. If your APIs are confusing with conflicting paths, you're compromising your chances of being useful to agents.
Levie cites Y Combinator's Jared Friedman putting developers on notice: even the best developer tools still don't let you sign up for an account via API. In the Claude Code age, that means your tool is effectively invisible to the agents that are increasingly doing the work.
At Box, Levie says they've been "combing over every aspect of our API to figure out what breaks down in a world of agents" — applying the same rigor to API design that previously went only into UX design.
Why This Matters Now
This isn't a prediction about 2030. It's happening right now. Coding agents like Claude Code, Devin, Codex, Cursor, and Replit already have their own sandboxed compute environments, file systems, long-term memory, and the ability to call APIs directly. They're not chatbots anymore — they're autonomous workers.
And it's gone beyond coding. Claude Cowork, Perplexity Computer, Manus, and OpenClaw have pushed agents into all areas of knowledge work. OpenClaw runs 24/7 in its own persistent environment. These agents review contracts, handle customer support, audit financials, generate code, and create presentations.
Levie's math is sobering: enterprises could easily have 100X to 1,000X more agents than people. Trillions of agents across the economy, all needing software to work with.
The New Infrastructure Stack
If agents are the new users, they need their own infrastructure:
Compute: E2B, Daytona, Modal, and Cloudflare are building sandboxed environments for agents. The next hyperscaler won't serve human applications — it'll serve agents.
Identity: Agents need their own identities, authentication, and access controls. Agentmail is already providing mailboxes for agents to have persistent email.
Search: Parallel, Exa, and others are rebuilding web search specifically for agent crawlers, not human browsers.
Payments: Agents will need their own budgets via Stripe or Coinbase wallets. Microtransactions might finally have a real use case — agents paying for paywalled tools and information.
Security and Governance: When agents access sensitive data in regulated industries like pharma or banking, companies need to govern and audit everything those agents did.
The Business Model Problem
Here's what should keep SaaS executives up at night: per-seat pricing doesn't work when your "users" are agents. With a few lines of text, an agent might do hours of human-equivalent work, only exposing the final output to the end user. That breaks the economics of seat-based software.
Levie says every tool that wants to survive needs consumption or volume-based pricing built in — including letting agents make their own payments for services they use.
The Bottom Line
Levie's essay is a wake-up call for every software company: the era of building for human users is ending. The companies that win the next decade will be the ones that make their software trivially easy for AI agents to discover, sign up for, authenticate into, and operate at scale. If your software still requires a human to click through a UI, you're building for the past. The future has APIs, and it has trillions of agents ready to use them.