With Vercel’s new Eve framework, the file system is the base upon which agents are built, and Vercel’s proprietary cloud services power the backend.
This architecture aims to simplify agent design and use, according to the company.
“Eve is designed around the idea that building an agent should mean defining what it does without assembling all of the pieces that it needs to run in production,” a group of the Vercel engineers wrote in an introductory blog post.
Vercel is positioning Eve, an npm package, as the next Next.js, potentially standardizing agent management much like the JavaScript framework standardized web development. “Agents today are where the web was before frameworks, with everyone hand-rolling the same plumbing and nothing carrying over to the next one,” the engineers wrote.
Eve can be freely downloaded – it is under an Apache 2.0 license – but the additional Vercel features, including orchestration, state management, and code execution sandboxes, will cost $20 a month. Enterprise rates are also available.
File System as Interface
“The filesystem is the authoring interface,” the GitHub page for Eve states.
Giving agents their own spaces on a file system seems to be gaining industry consensus, especially as an alternative to code-first approaches that require users to write software objects and API calls to define agent behavior.
“A filesystem isn’t a complete memory architecture […] but it’s an unusually good substrate for the part of memory agents struggle with most: durable, inspectable, revisable working context,” wrote Box head of Developer Relations Carter Rabasa, in an X post entitled “File Systems are the New Primitives for AI Agents.”
“Files give the model names, paths, hierarchy, timestamps, permissions, diffs, and conventions it already knows how to reason about.”
Some vendor-agnostic frameworks already use file system primitives as scratch pads, such as OpenClaw, LangGraph, and Mastra.
Eve takes the next step, using the file system directory structure as a workflow router.
Convention over Configuration
Borrowing the Next.js approach, Eve defines an “agent” simply as a set of directories. To build an agent, the user creates a scaffolding of directories, and places instructional files in the right folders, using Markdown.
The agent folder will include the following files and subdirectories, and be populated by the user:
- agent.ts: model and runtime configuration file, including the specification of which model to use
- instructions.md: the always-on system prompt and set of instructions that define the agent’s behavior.
- tools/: A folder of typed functions the user creates that the model can call (i.e. “get_weather.ts”)
- skills/: A folder for procedures loaded on demand (“plan_a_trip.md”)
- channels/: A folder for message channels (“slack.ts”)
- schedules/: A folder for recurring cron jobs (“weekly_recap.ts”)
To invoke an agent, the framework scans the directory, compiles any components found within, and then generates the execution loop that runs in the background.
Every agent gets its own sandbox, an isolated environment to which it is restricted. Set properly, the agent can’t accidentally delete a production database. Human-in-the-loop approvals can be added to a workflow by setting a single Boolean flag. Doing so can, for instance, prevent surprising bills when an agent consumes more data than necessary.
The user then can add capabilities to an agent by giving it either tools, written in TypeScript, or skills, which are instructions written in Markdown. Both tools and skills can also be written by the agents themselves. The agents have control over file system operations in their chosen directories, so they are able to write, read and search through directories.
Another advantage to this file system-first approach is that agents can be managed through existing Git and standard continuous development build pipelines.
The Vercel Cloud
The downside of Eve is that, for production usage, it must run on the proprietary Vercel cloud platform. It can be run on a laptop, but doing so causes the user to miss out on Eve’s major features (An enthusiast could swap out the Vercel AI Gateway for OpenRouter, but documentation is scarce and only time will tell if an open source community will rally behind making a truly vendor-agnostic version of Eve).
But there are upsides to moving your working directory of agents to the Vercel cloud. There, Eve maintains the state of each agent session, allowing the agent to resume work should it crash.
Agents can also access external resources – such as backends, data, and other third-party services – by tapping into Vercel Connect, which points to available MCP servers and OpenAPI-based APIs. Out of the box, Connect communicates with Slack, GitHub, Snowflake, Salesforce, Notion, and Linear. Agents can also delegate tasks to sub-agents.
Eve hands control of the tools to the model, brokering the authorization – the model never sees the connection’s URL or credentials.
Vercel initially built Eve to manage its own agents. The company’s engineers found that among the many different types of agents they built, “they all had the same needs, and the same structure kept emerging to meet them,” the engineers wrote.
Internally, Vercel has agents for data analysis, where users can ask questions about issues covered in existing documentation. The agent is restricted to the user’s permission level. Another agent handles sales leads, and one handles customer support ticket resolution.
