On June 17, 2026, Vercel open-sourced eve, a filesystem-first agent framework — and the same codebase that more than a hundred of Vercel’s own production agents run on. The pitch fits in one sentence: durable execution, sandboxed compute, human-in-the-loop, subagents, and evals — the wheels every agent team eventually reinvents — all ship in the box. You define what the agent does; the framework handles how it survives in production.
An Agent Is a Directory: Shape Written into the Filesystem
eve’s core design decision is that an agent is a directory, and the tree structure is the agent’s definition.
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 tasksIn agent.ts, defineAgent sets the model in one line, with provider fallbacks supported through AI Gateway; compaction, model options, and other optional fields are there when you need them. instructions.md is placed before every model call. Tools and skills need no registration: the filename and its position are the definition, everything loads at build time, and the filename becomes the tool name the model sees — no boilerplate. A skill file declares its own load timing in a description — “load before answering any revenue question,” for instance — so domain knowledge only enters the context when a relevant question appears. The feel is exactly Next.js for routing: the framework fixes the shape first, you fill in behavior.
Production Preloaded: Durable Sessions and Sandboxes
Every conversation is a durable workflow: each step is checkpointed, a session can pause for hours or days while waiting on an approval, and it survives crashes and deployments, resuming exactly where it stopped. This durability is built on the open-source Workflow SDK.
The sandbox posture matters more: agent-generated code is treated as untrusted, and every agent gets its own isolated environment for shell commands, scripts, and file I/O. The announcement’s own example is “break last week’s revenue down by region and chart it”: the agent writes analysis/by_region.py itself, runs it with bash in the sandbox, reports $2.1M for the Americas, $1.6M for Europe/Middle East/Africa, and $0.5M for APAC, and saves the chart to analysis/by_region.png. When a task needs a tool that does not exist yet, the agent can write one on the spot. The backend is an adapter design — Vercel Sandbox when deployed on Vercel, Docker, microsandbox, or just-bash for local development.
Human-in-the-loop compresses to a single field on a tool: needsApproval takes a condition — say, require sign-off only when a query scans more than 50GB. The agent pauses there indefinitely without consuming compute, then continues from the same point once approved.
The Human Boundary: Connections, Channels, and Subagents
Outbound integrations are connection files: one file points at an MCP server or an OpenAPI-compatible API, and the model never sees the URL or the credentials — OAuth is handled by Vercel Connect. At launch, Slack, GitHub, Snowflake, Salesforce, Notion, and Linear connect out of the box.
A channel is just a small adapter file: the HTTP API is on by default, Slack, Discord, Teams, Telegram, Twilio, GitHub, and Linear are built in, defineChannel supports custom ones, and channels can hand off to each other — the same agent can live on several surfaces at once.
Delegation runs through subagents: a subagent is a same-shaped directory under subagents/, with its own instructions, tools, and sandbox. The parent calls it just like a tool; the child starts with a clean context window and only the tools granted to it.
Observability and Verification: Traces, Evals, and Deploy Rhythm
Every run produces a trace listing the inputs and outputs of each model and tool call, plus the commands executed in the sandbox. The spans are standard OpenTelemetry, exporting to Braintrust, Raindrop, Arize, Honeycomb, Datadog, or Jaeger. Evals are files too: defineEval writes scored test suites, and eve eval runs them locally or pointed at a deployed app; wired into CI, it becomes a deploy gate.
Local development is one command — eve dev starts a dev server and a terminal UI. The key point is that the TUI is just a client: the agent serves the same structured events over HTTP, so curl, test scripts, or CI can drive it. Deployment claims zero provisioning: the agent is an ordinary Vercel project, vercel deploy puts it live, deploys do not interrupt running sessions, schedules deploy as Vercel Cron Jobs, and every commit gets a preview deployment with instant rollback.
Vercel’s Own Roster: A Hundred Agents Running (Vendor-Reported)
eve’s strongest argument is Vercel itself. Those hundred-plus agents began as separate projects on separate stacks, each with its own way of holding state, brokering credentials, and emitting logs — which, per the announcement, is where most teams find themselves after their second or third agent. Today they live in one monorepo, and because they share a shape, a hundred agents are built, observed, and upgraded with the same tools and conventions as one. The named cases (all numbers below are Vercel’s own account, with no third-party verification):
| Agent | Role | Reported number |
|---|---|---|
| d0 | Data analyst in Slack; every query scoped to the asker’s own permissions — it can never show you a table you could not already see | 30,000+ questions per month |
| Lead Agent | Runs the best rep’s playbook around the clock | About $5,000 a year, returns 32x |
| Athena | Built by RevOps in six weeks, no engineers | Pipeline coverage nearly doubled |
| Vertex | Around-the-clock support | Solves 92% of tickets on its own |
Beyond the table, two operational details are worth remembering: Lead Agent works every new lead the moment it comes in and follows up on its own, so none go cold overnight, with one engineer maintaining it part-time; Athena answers pipeline and forecast questions from Snowflake and Salesforce in plain language. One directional number backs the trend: a year ago, agents triggered less than 3% of deployments on Vercel; now they trigger around 29%, and the company expects half of all deployments to come from agents soon.
Limits, Trade-offs, and Builder Advice
The trade-offs come first: launch deployment supports Vercel only, with other platforms officially “on the way”; the framework integrates deeply with the Vercel stack (Sandbox, AI Gateway, Observability), so leaving Vercel means wiring your own adapters, and the announcement does not name a license — cite the repo for that. One more easy misreading: the announcement never says “AI SDK”; the only indirect signals are trace span names like ai.streamText and the provider fallbacks — reading eve as an AI SDK wrapper is inference, not a company statement.
Nor is it a no-code builder: tools stay TypeScript, knowledge stays Markdown, and what you learn is the directory convention. The concrete advice: eve ships as an npm package with development open at github.com/vercel/eve; the cheapest validation is to take one existing agent and rewrite it in this directory structure, and feel how much closer the remaining code sits to intent once “one file per ability” strips the boilerplate away.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.