Summary: eve was open-sourced by Vercel on June 17, 2026 — a filesystem-first agent framework, and the very code that more than a hundred production agents inside Vercel run on. The entire pitch is a single sentence long: durable execution, sandboxed compute, human-in-the-loop, subagents, and evals — the parts every agent team sooner or later rebuilds by hand — come installed from the start.
eve was open-sourced by Vercel on June 17, 2026 — a filesystem-first agent framework, and the very code that more than a hundred production agents inside Vercel run on. The entire pitch is a single sentence long: durable execution, sandboxed compute, human-in-the-loop, subagents, and evals — the parts every agent team sooner or later rebuilds by hand — come installed from the start. You describe what the agent should do; the framework takes care of how it stays alive in production.
One Directory, One Agent: the Definition Lives in the Filesystem
eve stakes its core design on one idea: an agent has no separate definition file — the directory is the agent, and its layout is what the agent is.
agent/
├── agent.ts # defineAgent: one line picks the model
├── instructions.md # system prompt placed before every call
├── tools/ # tools: one TypeScript file each
├── skills/ # domain knowledge: one Markdown file each
├── subagents/ # delegable child agents
├── channels/ # channel adapters
└── schedules/ # scheduled tasksThe model choice lives in agent.ts, where one line of defineAgent settles it; provider fallbacks are underwritten by AI Gateway, and optional fields such as compaction and model options stay off until needed. Ahead of every model call, instructions.md is inserted first. Tools and skills skip registration — where a file sits and what it is called completes the definition, loading happens at build time, and the tool name the model sees is the filename itself, no boilerplate anywhere. A skill file uses its description to state when it should load — “load before answering any revenue question,” for example — so the domain knowledge enters context only when a relevant question shows up. The approach is Next.js’s routing playbook: the framework pins the shape down, behavior is yours to fill in.
Production Gear in the Box: Durable Sessions and Sandboxes
A conversation in eve takes the form of a durable workflow: checkpoints land on every step, a session can sit paused for hours or even days while an approval is awaited, and it rides out crashes and deployments, resuming from the interrupted step exactly as it was. Underneath sits the open-source Workflow SDK.
The sandbox’s default stance deserves equal attention: agent-produced code is handled as untrusted across the board, each agent getting its own isolated environment for shell, scripts, and file I/O. The example the announcement gives is “break last week’s revenue down by region and chart it”: the agent writes analysis/by_region.py itself, pushes it through bash inside the sandbox, reports $2.1M for the Americas, $1.6M for Europe/Middle East/Africa, and $0.5M for APAC, and stores the finished chart as analysis/by_region.png. When a task lacks a tool that does not yet exist, the agent can write one on the spot. The backend follows an adapter design — Vercel Sandbox when deployed on Vercel, with Docker, microsandbox, or just-bash for local development.
Human-in-the-loop shrinks down to one field on a tool: needsApproval can carry a condition — for example, human sign-off is demanded only when a query scans more than 50GB. The agent halts there indefinitely and consumes no compute; once approved, it carries on from the same spot.
Where the Agent Meets the Outside: Connections, Channels, and Subagents
Outbound connections are managed through connection files: one file aims at one MCP server or one OpenAPI-compatible API; neither the URL nor the credentials reach the model’s view, and OAuth is handled by Vercel Connect. Slack, GitHub, Snowflake, Salesforce, Notion, and Linear all connect directly at launch.
A channel is, at bottom, one small adapter file: the HTTP API ships switched on, Slack, Discord, Teams, Telegram, Twilio, GitHub, and Linear are built in, custom ones go through defineChannel, and channels can hand off between themselves — one agent can thus live on several interfaces at once.
Delegation is the job of subagents: a subagent is another directory of identical shape under subagents/, equipped with its own instructions, tools, and sandbox. The parent calls it the way it calls any tool; the child sets out with a clean context window and carries only the tools granted to it.
Visible and Verifiable: Traces, Evals, and the Deploy Pipeline
Every run leaves behind a trace that itemizes the inputs and outputs of each model and tool call, along with the commands executed inside the sandbox; the spans are standard OpenTelemetry and can be exported to Braintrust, Raindrop, Arize, Honeycomb, Datadog, or Jaeger. Evals follow the same file-based pattern: defineEval writes them as scored test suites, and eve eval runs them locally or pointed straight at a deployed app; once CI is in the loop, the eval acts as a deploy gate.
Local development collapses into a single command — eve dev starts a dev server and a terminal UI together. What matters: the TUI is only ever a client — the agent serves the identical structured events over HTTP, so curl, test scripts, or CI can drive it just the same. Deployment is claimed to need zero provisioning: the agent is an ordinary Vercel project, vercel deploy puts it live, deploys leave running sessions untouched, schedules go out as Vercel Cron Jobs, and each commit receives a preview deployment with instant rollback.
The Scorecard Comes from Vercel Itself: A Hundred Agents Running (Vendor-Reported)
The most weighty evidence is Vercel’s own operation. Those hundred-plus agents used to be separately built projects on separate stacks, each keeping state, brokering credentials, and emitting logs its own way — per the announcement, the situation most teams land in once a second or third agent exists. They now share a single monorepo, and the shared shape means a hundred agents get built, observed, and upgraded with the same tools and conventions that serve one. A few cases are named (every figure below comes from Vercel’s own account, without third-party verification):
| Agent | Role | Reported number |
|---|---|---|
| d0 | Data analysis Q&A in Slack; queries are scoped to the asker’s permissions — a table you cannot see is one it cannot see either | 30,000+ questions per month |
| Lead Agent | Works the best rep’s playbook day and night | About $5,000 a year, returns 32x |
| Athena | Put together by RevOps in six weeks, no engineers involved | Pipeline coverage almost doubled |
| Vertex | Support that never clocks out | 92% of tickets resolved on its own |
Two details from outside the table deserve a note: Lead Agent takes up each new lead the second it arrives and follows through on its own, letting none sit overnight and go cold, and a year of running it takes one engineer part-time; Athena handles pipeline and forecast questions from Snowflake and Salesforce in natural language. A directional figure lends the trend further support: a year ago, agents accounted for less than 3% of Vercel deployments; the share is now roughly 29%, and the official expectation is that the halfway mark arrives soon.
Boundaries, Costs, and Advice for Builders
The costs come first: at launch, deployment supports Vercel alone, and other platforms get nothing more official than “on the way”; the framework is tied deeply into the Vercel ecosystem (Sandbox, AI Gateway, Observability), so life outside Vercel means wiring adapters yourself, and the announcement names no license — treat the repo as the authority when citing. One more spot invites misreading: the words AI SDK never appear in the announcement’s body; the only indirect signals are the naming of trace spans (such as ai.streamText) and the provider fallbacks — describing eve as an AI SDK wrapper is an inference, not an official statement.
A no-code builder it is not: tools remain TypeScript, knowledge remains Markdown, and the thing to learn is the directory convention. For builders, the advice is hands-on: eve is published as an npm package, with development carried on in the open at github.com/vercel/eve; the cheapest validation is to pick one existing agent and rewrite it in this directory structure, then feel how close the leftover code sits to intent once “one file per ability” has cleared the boilerplate out.