For builders and creators, the release of eve by Vercel signals a shift away from wrestling with complex agent plumbing. Instead of assembling infrastructure, developers can now define an agent simply by specifying its purpose. The framework is available as the npm package eve, licensed under Apache-2.0, and powers more than a hundred agents in Vercel’s own production environment today.
The core philosophy is straightforward: build the agent, not the container. Eve treats the filesystem as the primary interface for durable backend agents. You create an agent by establishing a directory on disk, which acts as the contract for that agent’s existence and capabilities.
The structure is intentionally minimal. A running agent requires just two files: one to define the underlying model and another to set the instructions. The model configuration is a single line, supporting provider fallbacks via AI Gateway, while the instructions.md file serves as the system prompt injected into every model call.
An agent is a directory
Vercel argues that agents require a consistent shape. Teams often reinvent the same architectural patterns to meet standard needs. Eve codifies this structure into a framework where every capability has a specific folder location.
This layout maps capabilities directly to paths, creating a clear contract for how an agent operates:
| Path | Role | Format |
|---|---|---|
agent.ts | The model it runs on, plus runtime config | TypeScript |
instructions.md | Who it is, prepended to every model call | Markdown |
tools/ | What it can do; filename becomes the tool name | TypeScript |
skills/ | What it knows; loaded only when the topic comes up | Markdown |
connections/ | Secure links to MCP servers and OpenAPI APIs | TypeScript |
sandbox/ | Optional override of the agent’s sandbox; seeds workspace files | Directory |
subagents/ | Specialist child agents it delegates to | Directory |
channels/ | Where it lives, like Slack or HTTP | TypeScript |
schedules/ | When it acts on its own, on a cron | TypeScript |
lib/ | Shared authored code used across the agent | TypeScript |
Expanding functionality is as simple as adding a file. Eve detects these additions at build time and wires them in automatically, eliminating the need for boilerplate registration code. For instance, a tool is defined by a single TypeScript file containing a Zod input schema, where the filename and its location within the tree constitute its full definition.
What ships in the box
Vercel describes eve as ‘batteries included,’ offering six production-ready capabilities out of the box:
- Durable execution: Every conversation functions as a durable workflow with checkpointed steps. Sessions can pause, survive crashes or deployments, and resume exactly where they left off, built on the open-source Workflow SDK.
- Sandboxed compute: Code generated by agents is treated as untrusted. Each agent operates within its own sandbox for shell commands, scripts, and file operations. The backend adapts to run on Vercel Sandbox in production or Docker, microsandbox, or just-bash locally.
- Human-in-the-loop approvals: Any action can be configured to require approval. The agent pauses indefinitely without consuming compute resources until a human authorises it, at which point execution continues seamlessly.
- Secure connections: Connections are files pointing to MCP servers or OpenAPI-compatible APIs. Eve brokers authentication, ensuring the model never sees raw URLs or credentials. At launch, agents can connect to Slack, GitHub, Snowflake, Salesforce, Notion, and Linear.
- Channels: A single agent instance serves every surface. The HTTP API is enabled by default, with integrations for Slack, Discord, Teams, Telegram, Twilio, GitHub, and Linear included. Channels can also hand off tasks to one another.
- Tracing and evals: Every run generates a trace using standard OpenTelemetry spans, exporting to Braintrust, Honeycomb, Datadog, or Jaeger. Evals consist of scored test suites runnable locally or integrated into CI pipelines.
Use cases, with real examples
Vercel has published six agents it runs internally using eve:
- d0, the data analyst: Its most-used internal tool, answering more than 30,000 questions a month. Every query is strictly scoped to the asker’s own permissions.
- Lead Agent, the autonomous SDR: It processes every new lead and follows up independently. Vercel estimates it costs about $5,000 a year but returns 32 times that value, maintained part-time by a single engineer.
- Athena, the sales cockpit: RevOps built this in six weeks without engineers. It answers pipeline questions from Snowflake and Salesforce in plain language.
- Vertex, the support engineer: It manages tickets across the help center, documentation, and Slack. Vercel reports it resolves 92% of tickets autonomously, escalating the remainder.
- draft0, the content agent: It runs a review pipeline that catches glaring issues before a human editor sees the final piece.
- V, the routing agent: Tasks arrive in Slack first and are routed by V to the specific agent capable of handling them.
eve versus a hand-rolled agent stack
Most teams currently assemble these components manually for each new agent. The table below contrasts this approach with what eve provides:
| Capability | Typical DIY stack | eve |
|---|---|---|
| Authoring | Custom loop, manual tool registration | Files in a directory, discovered at build |
| Durability | Bespoke state and retry handling | Checkpointed durable workflow per session |
| Code execution | Self-managed container or VM | Per-agent sandbox via swappable adapter |
| Approvals | Custom pause-and-resume logic | needsApproval field on any action |
| Surfaces | One integration per channel | One adapter file per channel |
| Observability | Stitched together from logs | OpenTelemetry traces and evals built in |
| Deploy | Provision infrastructure | vercel deploy, runs unchanged from local |
This comparison reflects eve’s documented capabilities. Specifics of other frameworks vary by version and setup.
Getting started
You can scaffold and start a new agent with a single command. It installs dependencies, scaffolds the project, and launches a dev server.
npx eve@latest init my-agenteve dev runs the agent locally with an interactive terminal UI. eve eval executes your test suites. eve build compiles inspectable artifacts under .eve/.
Since an eve agent is an ordinary Vercel project, vercel deploy ships it to production without modification. The sandbox environment swaps to Vercel Sandbox automatically, requiring no code changes.
Key Takeaways
- eve is Vercel’s open-source, Apache-2.0 agent framework, currently in public preview.
- An agent is defined as a directory of files, where each folder maps directly to one capability.
- Durable execution is built-in, allowing sessions to checkpoint and resume after interruptions.
- The framework includes secure connections to major tools and channels out of the box. Source Read original →
