Trillions of Agents: What Aaron Levie's Thesis Means for European SMBs

The Thesis: Software Will Be Built for Agents

Aaron Levie – CEO of Box and one of the loudest voices in enterprise software – recently put forward a thesis that's hard to ignore: In a world of trillions of agents, agents will be the primary users of software. Not humans. Agents.

His full post on X reads like a manifesto for the next era of enterprise software. The core message:

"Make something agents want." – Aaron Levie, riffing on Paul Graham's famous "Make something people want."

Levie argues that agents are no longer glorified chatbots. They have their own sandboxed compute environments, file systems, long-term memory, can write and execute code, call APIs directly – and they do it 24/7. What started as coding agents with Claude Code, Cursor, or Replit has now spread to every area of knowledge work.

What Does "Trillions of Agents" Mean?

Levie's math is straightforward: if every employee in a company has dozens of agents working on their behalf, a single enterprise quickly has 100x to 1,000x more agents than people. Scale that across the global economy and you get trillions of software agents working in parallel.

It sounds like science fiction. But the building blocks already exist:

• Coding Agents (Claude Code, Codex, Cursor) write and deploy code autonomously
• Knowledge Agents (Perplexity, Manus) research and synthesize information
• Workflow Agents (OpenClaw, Claude Cowork) execute complex multi-step tasks
• Infrastructure Agents run in sandboxed environments on E2B, Modal, or Cloudflare

So What Does This Mean for SMBs?

This is where it gets interesting – and where we deliberately diverge from Levie's Silicon Valley perspective.

1. API-First Is No Longer Optional

Levie says: "If you don't have an API for a feature, it might as well not exist." For SMBs, this means: every tool you use must have an API. Every single one. And not just on paper – one that agents can actually consume.

Our take: We see this daily. Companies buy software with "API access" in the marketing materials, but the API is incomplete, undocumented, or requires manual setup steps that no agent can handle. Audit your tool landscape: which of your systems are truly agent-ready?

2. MCP Is Becoming the Standard Interface

The Model Context Protocol (MCP) – developed by Anthropic – standardizes how agents communicate with external tools. Levie explicitly mentions MCP as critical infrastructure. And he's right: MCP is to agents what REST APIs were to web apps.

For SMBs, this means: your systems need MCP servers. monday.com, Make, n8n, Supabase – the tools we use daily – are already working on MCP integrations. Early adopters will have a structural advantage.

3. Agent Skills > Traditional Documentation

Levie talks about "skills that agents use." That's exactly our Agent Skills Framework: behavior-based workflow knowledge encapsulated in SKILL.md files. Not passive documentation – but executable instructions for agents.

The SMB advantage: You have domain-specific knowledge that no foundation model knows. Your SOPs, your industry logic, your process quirks. Anyone who structures this knowledge as Agent Skills makes their agents instantly more productive than any generic AI solution.

4. Local Infrastructure Becomes Strategic

Levie focuses on cloud infrastructure. But for European SMBs, there's an additional factor: data sovereignty. Not every agent can or should run through a US-based cloud.

That's exactly why we built Project KNUT: 52 GB VRAM, heterogeneous hardware, local inference at 54 t/s – without a single byte leaving your network. In a world of trillions of agents, the question "Where do my agents run?" becomes a strategic decision.

What Levie Gets Right – and What's Missing

✅ Right: Business Models Must Change

Per-seat pricing doesn't work in a world where an agent completes hours of human work in seconds. Consumption-based pricing will become the standard. For SMBs, that's good news: you pay for output, not licenses.

✅ Right: Security and Governance Become Critical

Agents need their own identities, their own permissions, their own audit trails. Levie rightly calls this one of the biggest challenges. In the regulated European market (GDPR, financial compliance, pharma regulations), it's even more complex.

⚠️ What's Missing: The Human in the Equation

Levie's post is tech-optimistic – and that's fine. But he leaves out a point that matters to us: AI First = People First. Agents don't replace teams. They amplify them. The question isn't "How many agents can I deploy?" but "How do I use agents so my team can focus on what humans do best?"

Empathy, creativity, relationship building, difficult conversations – these aren't features you call via API. And that's where the SMB advantage lies: smaller teams, shorter feedback loops, real relationships. Agents should amplify that advantage, not replace it.

Concrete Steps: What Should You Do Now?

Levie's article is a wake-up call. But wake-up calls need action. Here are five steps you can start this week:

1. Run an API audit: List every tool you use. Which ones have a complete API? Which don't? The ones without APIs are on the chopping block.

2. Evaluate MCP: Check which of your tools already offer MCP servers. Test the integration with Claude Code or another coding agent.

3. Write your first Agent Skills: Take a recurring process (onboarding, reporting, QA) and formulate it as a SKILL.md. Test whether an agent can execute it.

4. Clarify data sovereignty: Which data can go to the cloud? Which must stay local? Define zones before you let agents loose on sensitive data.

5. Involve your team: Agents will only create value if your team understands what they do and how to work with them. Invest in Agent Literacy.

Conclusion: The Future Belongs to Those Who Build Now

Aaron Levie is right: we're heading toward a world where agents are the primary users of software. But the future won't be decided in San Francisco – it will be built by the companies that set the course now.

For European SMBs, that means: don't wait. Don't start an innovation project first. Make your systems agent-ready now, write your first skills, and find out where agents create real impact.

We can help. Get in touch.

---

→ Claude – AI for Coding, Analysis & Autonomous Agents

Source: Aaron Levie on X – "Building for trillions of agents", published April 2026.