Most agent frameworks hand you a reasoning loop and wish you luck. Checkpoint logic, sandbox isolation, approval gates, observability - those are your problem. Every team building a production agent that runs for hours, waits on human decisions, integrates with external APIs, and survives server crashes has to assemble the same scaffolding before the agent does anything useful. Most agent frameworks supply the logic and leave all of that to the team.
Vercel's answer to that problem is Eve, an open-source TypeScript agent framework released June 17 at its Ship 26 conference in London.
The problem is genuine, even if the solution is opinionated.
What Eve actually ships
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. It comes with durable execution, sandboxed compute, human-in-the-loop approvals, subagents, and evals built in.
An Eve agent is a directory of files: instructions.md sets the system prompt, TypeScript files in agent/tools/ define callable tools, and Markdown files in agent/skills/ serve as on-demand playbooks.
Add a TypeScript file to tools/ and the model can call it. The filename becomes the tool name. No registration required.
Agents wait on people, call slow systems, and run for hours, days, or weeks. In Eve, every conversation is a durable workflow with each step checkpointed, so a session can pause, survive a crash or a deploy, and resume exactly where it stopped. This durability is built on the open-source Workflow SDK.
The sandbox story is equally direct. Code your agents write should be treated as untrusted. Eve keeps agent-generated code out of your application runtime entirely. Every agent gets its own sandbox - an isolated environment for shell commands, scripts, and file reads and writes - running in a separate security context from the harness that controls the agent. Locally that means Docker or microsandbox; in production it runs on Vercel Sandbox without a code change.
Skills are Markdown playbooks loaded when they are relevant. The agent gets focused guidance without carrying it in every prompt. That is just good context management - only the relevant playbook rides along on a given tool call, rather than bloating every request with a monolithic system prompt.
The Vercel-specific catch
This is where the honest accounting matters. Eve's production capabilities - durable execution via Vercel Workflow, sandboxed compute via Vercel Sandbox, model routing via AI Gateway, credential brokering via Vercel Connect, observability in the Agent Runs dashboard - are each built on Vercel's platform. Independent analysts reviewing the launch noted that porting Eve to a non-Vercel runtime requires replacing each of these adapters.
This is the same dynamic developers already documented with Next.js, where features like Incremental Static Regeneration and Edge Middleware work seamlessly on Vercel but require significant rebuilding on other hosts.
The framework's adapter pattern technically permits replacement, but the path is not frictionless. If you are evaluating Eve against something like Mastra - a Y Combinator-backed TypeScript framework that reached version 1.0 in January 2026, designed to run on any host
- or LangGraph with its own checkpointing primitives, the honest question is whether the zero-config deploy experience is worth the platform dependency.
The framework is model-agnostic - it routes requests through Vercel's AI Gateway and supports any model provider without changing orchestration logic. Model lock-in is not the issue. Infra lock-in is.
What Vercel says it runs on Eve
This part is harder to dismiss than the pitch. Vercel published six agents it runs internally on Eve. Its most-used internal tool, d0, handles more than 30,000 questions a month. Every query is scoped to the asker's own permissions.
Vertex, the support engineer, handles tickets across the help center, docs, and Slack. Vercel reports it solves 92% of tickets on its own and escalates the rest.
That is the kind of production evidence that separates a framework announcement from a demo. The fact that Vercel itself is the customer confirms the framework had to survive real operational load before shipping.
CEO Guillermo Rauch said agents now trigger more than half of all commits on Vercel's platform, up from under three percent six months ago. That claim deserves skepticism - "commits" and "deployments" can be defined narrowly - but the direction is consistent with what teams building on coding agents are reporting more broadly.
Is this genuinely new, or incremental?
The filesystem-first structure (one directory, one agent) is a design choice, not a technical breakthrough. Durable execution, sandboxing, and human-in-the-loop approvals all exist in other frameworks. What Eve packages differently is that the framework stops hand-rolling the same production plumbing for every agent and ships durable execution, sandboxes, approvals, and evals as defaults, not weekend projects.
The genuinely new element is the channel abstraction.
Sessions move between channels - a question asked in Slack can continue on the web, and an incident webhook can open an investigation thread in Slack. At launch: HTTP API, Slack, Discord, Teams, Telegram, Twilio, GitHub, Linear, plus a defineChannel for custom surfaces.
For a workplace context, that matters. An agent that can start a conversation in Slack, receive an approval click from the same Slack message, then resume a long-running task - without duplicating any state - is a meaningfully different interaction model than re-triggering an agent from a form each time.
If you already ship Next.js on Vercel, Eve is the lowest-friction path from "agent demo in a notebook" to "agent with approvals, schedules, and a Slack channel in production." The filesystem layout also plays nicely with coding agents: Claude Code or Cursor can add a tool by creating a file in agent/tools/ without touching framework internals.
The honest verdict
Eve earns its announcement. The production defaults are real, the internal case studies are credible, and the channel model is more composable than most agent surfaces today. The tradeoff is not hidden: this is infrastructure for teams already on Vercel, or teams willing to be. For everyone else, you are back to assembling the pieces yourself - or choosing a platform-neutral framework and accepting more setup work.
A teammate like Beagle, living in Slack and Teams, would slot into Eve's channel model without much ceremony. But the broader point applies to any agent being evaluated for workplace deployment: check whether the durability and approval story is built into the framework or left to you. With Eve, at least, the answer is clear up front.