Vercel released eve on June 18, an open-source agent framework designed to eliminate the infrastructure work that currently precedes any production agent deployment. The announcement, published on the Vercel blog by Shar Dara, comes with a public preview and a CLI wizard that scaffolds a running agent in under a minute.
The core design decision in eve is that an agent is a directory. Each file in the tree describes one responsibility: agent.ts sets the model, instructions.md serves as the system prompt, and files under tools/, skills/, channels/, and schedules/ extend the agent’s capabilities without requiring any wiring code. Vercel’s position is that the same structural pattern kept emerging across every internal agent the company built, so they codified it into a framework rather than continuing to repeat it.
Production concerns ship with the framework rather than on top of it. Every conversation runs as a durable workflow with each step checkpointed, meaning a session can survive a crash or a mid-task deploy and resume from the exact stopping point. Agent-generated code runs inside an isolated sandbox, separate from the application runtime. Any tool can be flagged to require human approval before execution, and the agent pauses without consuming compute until that approval arrives. Traces are emitted as standard OpenTelemetry spans and export to Braintrust, Honeycomb, Datadog, or Jaeger.
The connections model is worth noting for teams already invested in MCP infrastructure. A connection in eve is a file that points at an MCP server or any API with an OpenAPI document. The framework discovers available tools, brokers authentication, and keeps credentials out of the model’s context window. At launch, eve ships with built-in connections for Slack, GitHub, Snowflake, Salesforce, Notion, and Linear.
Vercel claims to run more than a hundred agents in production on eve internally. The numbers cited in the announcement are specific enough to be useful as reference points. The company’s autonomous SDR agent costs roughly $5,000 per year to operate and the company states it returns 32 times that figure. A support agent called Vertex resolves 92 percent of tickets without human escalation. A data analyst agent named d0 handles more than 30,000 questions per month from the Snowflake warehouse, with every query scoped to the requesting user’s existing permissions. Vercel built its RevOps forecasting tool, Athena, in six weeks without engineering involvement.
The comparison Vercel reaches for is Next.js ending the era of hand-rolled web server plumbing. The analogy is structurally honest in one respect: the framework earns its abstractions after enough people have built the same thing the hard way, and agent infrastructure is now old enough to standardize. The question every team will ask is whether the deploy-to-Vercel constraint fits their architecture. The announcement says support for other platforms is coming, but eve currently deploys to Vercel exclusively. For teams on AWS, GCP, or a self-hosted stack, the local dev experience and the open-source runtime are available today; the production path is not yet provider-agnostic.
Evals are a structural feature rather than an afterthought. Test suites live as files alongside the agent, run locally or against a deployed instance, and can gate CI deployments. Every commit also gets a preview deployment that carries the agent’s channels with it, so a team can interact with a staging version of their Slack bot before it replaces the one in production.
The agent-deployment share on Vercel’s own platform has moved from under 3 percent of deployments a year ago to roughly 29 percent today, and the company projects that agents will account for half of all deployments soon. That trajectory is worth treating as a directional signal, even accounting for the fact that Vercel has strong incentive to frame the number favorably.
Teams currently evaluating agent frameworks should run eve’s init command against a real internal use case before committing to a stack, particularly to assess whether the directory-as-agent model and the Vercel deployment requirement fit their production constraints.
Source: Vercel blog, published June 18, 2026, authored by Shar Dara.
