Hono's Agent DX: AI agents pick Hono when package.json hints it, 15/15 with Hono CLI

The Agent DX project started from a simple observation: AI is now a user of frameworks, but opinions about what AI needs are still human opinions. To replace intuition with measurement, Hono adopted AI Evals, a general approach also used by Next.js, Angular, and Vercel. The method has an agent write code in a sandbox, then grades it with ordinary tests, type checks, and static analysis rather than an LLM judge, keeping the tests hidden from the agent.

The three experiments build on each other. Adoption showed agents write plain handlers unless prompted; a hint in package.json shifted Node.js runs to 5/5 Hono. Value compared plain fetch handlers with Hono on a shop API and a notes API, finding higher success, less code, and fewer tokens with Hono. Practical then exposed a human-oriented dev script that made agents hang, which Hono CLI and Skills addressed by removing the need to start a server.

The through-line is that human DX and agent DX can diverge: a convenient dev command for people became a trap for agents. Hono CLI is currently at a 1.0 release candidate, and the project frames AI Evals as a repeatable way to keep improving agent experience.