Vercel's eve: What It Unlocks for Production AI Agents
An AI agent becomes useful in production when it can keep track of unfinished work. Customer replies arrive hours later. Staff decisions interrupt the flow, and external systems return results on their own schedule. Deployments land before a conversation is done. Vercel's eve is a serious attempt to give those ordinary interruptions a dependable home.
The framework brings durable sessions and approval flows together with sandboxed compute, channel adapters, connections, scheduled work, delegated agents, tracing, and evaluation. Each part has existed elsewhere. eve packages them around a clear way to define and operate an agent.
At a glance: eve is an open-source TypeScript framework for agents that need to pause, resume, use tools, work across channels, and remain observable in production. Its strongest fit is a product built on Vercel that needs long-running agent work and human oversight.
How eve is structured
eve is an open-source, filesystem-first TypeScript framework. An agent is represented by a directory.
Permanent instructions live in Markdown, and tools live in TypeScript. Skills, channels, schedules, connections, hooks, sandboxes, and subagents each have a conventional place in the project.
The convention makes an agent easier to inspect. A reviewer can see its role and capabilities alongside the places it runs and the work it performs on a schedule. Changes can move through Git, preview deployments, evaluations, and rollbacks as part of the application release process.
Vercel describes eve as “like Next.js for agents.” The analogy is useful within limits. eve supplies structure and runtime conventions around an agent. Teams choose the model, write the instructions, define the tools, and decide where human judgment belongs.
What the framework unlocks
Capable engineering teams can build each part of this stack independently. eve's value comes from placing the parts in one runtime with shared conventions.
| Capability | What it enables | Why it matters to users |
|---|---|---|
| Filesystem-first agents | Instructions, tools, skills, channels, and schedules can be reviewed together | Agent behavior can improve through the same disciplined release process as the product |
| Durable sessions | Work can pause, survive a deployment, and resume later | A request can remain active during a reply or review |
| Human approvals | An agent can wait for a person before continuing | Staff can supervise important work and keep the same conversation |
| Sandboxed compute | File and code tasks can run in an isolated environment | Agents can prepare more useful work in a separate environment |
| Channels | One agent can serve web, messaging, API, cron, and other surfaces | Customers can receive a more consistent experience wherever they start |
| Connections | Agents can work with external services and MCP tools | A conversation can lead toward a practical next step |
| Subagents and schedules | Specialized or recurring work can run with focused instructions | Agents can support deeper tasks and timely follow-up |
| Traces and evals | Teams can inspect runs and test changes before deployment | Reliability can improve through evidence |
Durable execution keeps work moving
Customer work rarely follows the timing of one request and one response. A question arrives after hours, then the customer replies the next morning. An appointment request waits for a staff decision.
According to Vercel's eve launch post, every conversation runs as a durable workflow with checkpointed steps. A parked session can survive a crash or deployment. Later, it resumes from stored state with its place in the conversation intact.
Human approval becomes a natural pause
eve allows a tool to request approval and pause the session. After a person decides, the session continues. The approval can appear inside the channel where the work is already happening.
A staff member gets the proposed action together with the surrounding context. The customer can keep the same conversation, and the team can define which decisions deserve review.
Sandboxes support work beyond conversation
Agents increasingly work with files, calculations, data transformations, and drafts. eve gives an agent an isolated sandbox for file and shell tasks. Vercel-hosted deployments can use Vercel Sandbox, and local or self-hosted adapters are available.
That environment lets an agent prepare work for review. A support agent could turn supplied information into a document or run a bounded calculation before a staff member takes the next action.
The boundary depends on configuration. Vercel's harness documentation recommends reviewing and restricting tools that reach the filesystem, network, shell, or sensitive data.
One agent can meet users in more places
An eve agent can serve multiple channels, connect to external tools, delegate focused work, and run on a schedule. A role defined once can answer on the web, join a messaging channel, wait for a person, and follow up later through an authorized service.
The framework also supports traces and evaluations. On Vercel, agent runs can appear in Observability with model calls, tool calls, and sandbox commands shown in sequence. Teams get a concrete record when they investigate a successful run or a failure.
The honest caveats
eve remains in beta, so its APIs and behavior remain subject to change before general availability. The active changelog shows a framework moving quickly. Early adopters should plan for regular upgrades.
The shortest production path runs through Vercel's own stack, including Workflows, Sandbox, AI Gateway, Connect, previews, and Observability. eve carries an Apache-2.0 license and documents self-hosting. Teams operating it elsewhere carry more infrastructure work themselves.
Vercel also reports extensive internal use. The company says it runs more than 100 production agents on eve across data, sales, support, content, and routing. That figure shows substantial first-party use. An independent benchmark would answer different questions about portability and performance across teams.
Adopters should test eve against their own operational needs, especially recovery behavior and approval flows. Tool boundaries deserve the same scrutiny as any other production integration.
Why Agent One adopted eve
Disclosure: Agent One has adopted eve. The independent framework review above covers the technology. This section covers what the adoption enables for our users.
Agent One gives a business one trained AI role that answers customers, qualifies requests, prepares the next step, involves staff when needed, and keeps the experience moving. Those conversations often span more than one sitting and depend on a person at key moments.
Agent One entered this adoption with persistent conversation records and a browsable history. It already had a live inbox and customer channels across web chat, Slack, and WhatsApp. Customers could ask for a person. Staff used the inbox to take ownership. When ready, they returned the conversation to AI.
eve gives that existing product a durable foundation for longer work. It also gives our team shared conventions for approvals, channels, isolated preparation tasks, and evaluation.
The handoff Agent One already supported
Human handoff was useful in Agent One before eve. A customer could ask for a person and the business received the conversation in its inbox. The team saw the transcript and could reply from the same interface. Takeover controls kept automated responses out of the way until staff returned the conversation to AI.
That experience was shaped like live chat. It worked best when a teammate replied quickly and the visitor kept the conversation open. We later improved conversation restoration for returning visitors. The customer's reply path stayed tied to the right active conversation.
Real support regularly spans lunch breaks and closed tabs. Some requests wait through a shift change or until the next morning. A saved transcript gave staff useful context during those gaps, and the eve adoption lets the customer-facing conversation remain active across them.
eve lets the handoff outlive the chat window
On an eve-enabled conversation, Agent One treats the handoff as ongoing work attached to the same customer request. A staff reply can arrive after the customer has left the page and remain available in that conversation. When the customer returns to it, the response is waiting.
Customers can keep adding details after a person takes over. Their messages stay in the thread for the team, and automated replies remain paused during the handoff. Staff decide when Agent One resumes, with the recent human exchange available as context.
For an eve-enabled conversation, full handoff support means:
- A customer can request a person and leave the chat window.
- The request remains visible to staff until the team claims or resolves it.
- A staff reply stays attached to the same conversation through the supported channel.
- The customer can keep adding context as a person handles the request.
- Agent One resumes after staff return control, carrying the recent exchange forward.
Full support here describes continuity across the handoff lifecycle. Staff response time and channel availability vary.
What eve extends in Agent One
The adoption extends capabilities that were already part of Agent One and adds work that became practical through eve.
| Customer need | Agent One already supported | The eve-enabled extension |
|---|---|---|
| Return to a conversation | Saved transcripts, conversation history, and restoration for returning visitors | A durable session keeps its place through a longer pause or deployment |
| Reach a person | An inbox, staff takeover, human replies, and a return-to-AI control | The handoff remains open after the live window closes, and recent human context follows the return to Agent One |
| Review an appointment | Supported scheduling tools could complete a booking during a live request | The appointment request can wait for authorized review and continue in the same conversation after staff decide |
| Use familiar channels | Customers could reach Agent One through web chat, Slack, and WhatsApp | Each supported channel can keep its own conversation intact under a shared durable operating model |
| Prepare the next step | Connected tools helped Agent One answer and act during a request | A bounded task such as a non-binding quote draft can run in an isolated workspace and return for staff review |
Agent One keeps a conversation intact within each supported channel. Verified linking determines when histories across channels belong to the same customer.
Approvals can wait inside the conversation
Earlier Agent One tools could complete supported actions during a live request. On the eve-enabled path, an appointment request can pause for the authorized staff member and continue after that person decides. The customer keeps the conversation where the request began.
The same approval pattern gives staff a clear place in the work. Agent One can prepare the next step and preserve the surrounding context until a person is ready to review it.
Useful work can be prepared within clear boundaries
Using eve's sandboxed-compute model, Agent One can turn provided inputs into a non-binding quote draft in an isolated workspace. Staff review the draft before it becomes a customer-facing decision.
A measured rollout
These changes apply to eve-enabled agents. Existing Agent One experiences remain on their current runtime as each migration is validated, and earlier conversations stay on the runtime where they began. Every eve activation passes readiness checks before it serves customers, keeping established experiences stable as the durable path expands.
The framework supports more than Agent One exposes today. We release each capability after it has a clear operating experience for the business.
What this unlocks for users
A reply can wait for the customer
A customer can leave after requesting a person. Staff can answer when they are available, and the response remains attached to that conversation for the customer's return.
Staff can take over and keep the thread
The inbox gives staff the request and the relevant conversation history. Customers can add another detail during the handoff, and the team sees it in the same place. When staff return control to Agent One, the recent human exchange remains available.
After-hours requests can stay open
Agent One can receive a request when staff are unavailable and gather the context the team will need. The work remains ready for the next person, so the customer receives an accurate account of where the request stands.
An approval can continue the same request
An appointment or another supported action can wait for staff review. After staff decide, the result returns to the conversation where the customer asked for it.
Each channel can keep its own continuity
Our first eve-enabled channel work spans web chat, Slack, and WhatsApp. The shared operating model helps the same role keep a consistent handoff standard and approval flow in each one.
Product work can focus on the experience
A shared agent runtime frees more engineering time for the parts users directly encounter. That includes easier setup and clearer controls, along with useful handoffs and a simpler view of work awaiting attention.
Where the adoption stands today
Agent One has adopted eve for its customer-operations runtime. Our first work on that path covers web chat, Slack, and WhatsApp experiences.
eve supports more channels, connections, schedules, subagents, and evaluation patterns than Agent One exposes today. We will release those capabilities according to the customer outcome they improve. A new channel earns its place when customers receive consistent help there. A schedule earns its place when timely follow-up improves the experience.
The verdict
eve makes a strong case for teams building agents that handle durable work across several surfaces. Its integrated approach to approvals and isolated execution addresses parts of production agent work that usually require substantial infrastructure.
The beta label and Vercel-centered production path are material tradeoffs. Teams already invested in Vercel get the clearest route to value, and teams elsewhere should price in the extra operational work. The framework's rapid development also calls for active maintenance.
For Agent One users, the adoption should show up as continuity. Conversations can retain context through pauses, and staff keep decision authority. As channel support expands, the same trained role can remain recognizable wherever a customer starts.
If you are defining what an AI operator should own, start with what an AI front desk should do. Then define the limits with our guide to human escalation for AI agents. When the role and boundaries are clear, you can build the first version in Agent One.