Setup
For two years, “AI-first” meant adding a chat box. That framing is now wrong. When Andrej Karpathy posted “Build. For. Agents.” in February 2026, he wasn’t describing a future state — he was describing a gap that already exists. Products built for human eyes and human hands are structurally inaccessible to the agents that increasingly do the work. The mismatch is no longer theoretical.
What the Shift Means
Karpathy’s argument is specific: if your product can’t be reached by a CLI, an MCP endpoint, a structured Skill, or plain Markdown output, it doesn’t exist for agents. The 2026 user is often not a person — it’s a loop. An agent calling your API, parsing your docs, chaining your tool with three others. Products that require a GUI, a login flow, or an AJAX-heavy interface drop out of that loop entirely. They become invisible.
How It Works in Practice
The architectural path is straightforward. MCP (Model Context Protocol) gives agents a typed interface to call your product’s capabilities. Skills wrap those capabilities with context so agents know when and how to use them. Markdown output means results that agents can read and pass downstream without a parser. CLI access means no browser dependency. These aren’t new technologies — they’re the same primitives that made Unix composable. The insight is applying them to the agent-integration layer deliberately, not as an afterthought.
Why It Matters
- SaaS companies that built for human workflows now face a distribution problem: their tools won’t appear in agent-composed stacks unless they add MCP or API-first access in 2026, not later
- Developer tools with strong CLI support (Cursor, GitHub, Vercel) gain compounding advantage as agents default to the tools they can already reach
- Startups building for “AI-native” workflows have a narrow window to define what agent-accessible infrastructure looks like before incumbents retrofit it