eve treats AI agents like conventional file trees

What it does

eve is a TypeScript framework from Vercel for building AI agents. It enforces a conventional directory structure—instructions.md, tools/, skills/, channels/, and schedules/ inside an agent/ folder—so an agent’s behavior is visible in the filesystem rather than buried in code. The model, typed tool definitions, and runtime configuration are all optional files that the framework picks up by convention.

The interesting bit

The framework treats the filesystem as the primary authoring interface, which is a deliberate bet against the usual kitchen-sink entrypoint. In a slightly recursive twist, the documentation ships inside the npm package so coding agents can read it locally from node_modules/eve/docs.

Key highlights

• Filesystem-first layout: prompts live in markdown files, tools in TypeScript, schedules as cron files.
• Supports typed tool definitions with Zod schemas and pluggable models.
• Includes channels for HTTP, Slack, and Discord out of the box.
• Ships its own documentation inside the eve package for local agent consumption.
• Explicitly beta; APIs and behavior may change before general availability.

Caveats

• The README promises “durable” agents but never explains whether that means persistence, fault tolerance, or simply long-running tasks.
• The framework is in public beta and subject to Vercel’s beta terms, meaning APIs and behavior may shift.

Verdict

Developers who believe an agent’s behavior should be discoverable by reading a folder tree will appreciate the convention-over-configuration approach; those who prefer explicit wiring in a single entrypoint may find the invisible plumbing frustrating.