Vercel Open-Sourced Eve — an Agent Framework Where Each Agent Is Just a Directory of Files

The News

Vercel open-sourced eve, a framework for building AI agents, under Apache 2.0. The pitch is structural, not conceptual: an eve agent isn't a class or a config blob — it's a directory. Drop a file in the right folder and eve wires it in automatically, no registration boilerplate:

• agent.ts — model configuration
• instructions.md — the system prompt
• tools/ — executable capabilities (the filename becomes the tool name)
• skills/ — knowledge loaded contextually, not always crammed into context
• connections/ — brokered links to external APIs and MCP servers
• channels/ — where the agent is reachable (Slack, Discord, HTTP, GitHub, Linear...)
• schedules/ — cron-triggered autonomous runs
• subagents/ — delegated specialists

If you've been building CLAUDE.md-driven workflows, subagent teams, or scheduled routines in this Academy's agentic modules, this structure will look familiar — it's the same instinct (separate the what from the how, make capabilities discoverable as files) formalized into a shipping framework.

The Technical Substance

Three pieces are worth understanding even if you never touch eve directly:

Durable execution. Every conversation is a checkpointed workflow, running on Vercel's open-source Workflow SDK. A session can pause, survive a crash or a redeploy, and resume exactly where it left off — instead of losing hours of agent state because a process restarted.

Sandboxed, untrusted code. Agent-generated code runs in an isolated sandbox (Vercel Sandbox in production, Docker or microsandbox locally) — a separate security context from the harness controlling the agent. That's the same principle last week's Grok Build leak violated: agent output and agent-with-your-credentials should never share a process.

Brokered connections, human approval gates. Connections to Slack, GitHub, Snowflake, Salesforce, Notion, and Linear are files that reference MCP servers or OpenAPI specs — eve brokers the auth, and the model never sees the raw URL or credentials. Any action can be marked needsApproval, which pauses the agent indefinitely, without burning compute, until a human signs off.

Why It Matters for You

None of this is exotic if you're already thinking in agent-oriented terms. What eve demonstrates is that these patterns — durable state, sandboxing by default, brokered secrets, human-in-the-loop gates — are becoming the baseline expectation for production agent frameworks, not advanced options. If you're designing your own agent tooling, steal the shape: files over config, sandboxes over shared processes, approval gates over blind autonomy.

---

Go deeper: Chapter 7 of the Vibe Coding Ebook covers Claude Code agent teams, Cursor Automations, and scheduled task patterns — the same territory eve's subagents/ and schedules/ directories formalize.

Read more on agent-driven dev: EndOfCoding tracks agent framework releases like this as they ship.

Hiring in this space: LLMHire is tracking a rise in "Agent Infrastructure Engineer" and "AI Platform Engineer" listings as teams like Vercel's ship production agent frameworks.

Source: Vercel, "Introducing eve" (June 2026); github.com/vercel/eve (Apache 2.0).