Vercel launches eve 1.0, AI agent framework with native React and Next.js integration

- Reading
- 7 min
- Tables
- 1
- Code
- 1
After months in public testing, eve reaches version 1.0 with native integrations for Next.js, SvelteKit, and other frameworks, promising response turns up to 7.3 times faster and significantly lower infrastructure cost.
Vercel announced eve 1.0, the stable version of an open-source framework for building production-ready AI agents. The project came out of the public preview opened in June 2026 after more than 230 releases and thousands of pull requests merged by the team and early adopters, according to the company's changelog. For those building front-ends with React and Next.js, the part that matters isn't just "another agent framework": it's the native integration that eve now offers for putting these agents inside the product's own interface.
The harness that hides the agent's infrastructure
The central concept of eve is what Vercel calls the harness: the loop that calls models, runs tools, and delegates work to subagents. Whoever builds the agent chooses the model, adds or swaps tools, edits instructions, and decides which actions require human approval before running. The Workflow SDK, also open source, saves execution progress at every step, so an agent can resume from the last completed point after a crash, without restarting the conversation from scratch.
Multiplayer sessions and agent-to-agent calls already come built into the harness. This matters because it's exactly the kind of state (who approved what, at what point execution stopped, who is in the session) that historically forced front-end teams to build their own state machine on top of a raw streaming API.
Where the UI comes in: native integrations with Next.js and friends
eve 1.0 expands native integrations to Next.js, Nuxt, SvelteKit, TanStack Start, and Vite. These integrations build the agent alongside the application and already set up the necessary routes; on Vercel, you can deploy the agent and the site as a single project or keep them separate.
The practical point for those working with React is the client library: it exposes tool calls, approval requests, and agent questions in a way that's consistent across frameworks. With this session data, you can stream responses in a chat, embed the agent on any page, or display tool calls in your own components, instead of relying on a generic chat widget.
For those who just want to get started from scratch, eve add channel/web sets up a complete chat using Vercel's own AI Elements components. For those who already have a UI in place, the way forward is to plug eve's React, Vue, or Svelte bindings into what already exists. In practice, this eliminates much of the work of manually rebuilding "waiting for approval", "tool running", and "response streaming in" states that today tend to vary from project to project.
Multi-tenant agents right in the UI
eve 1.0 also formalizes support for multi-tenant agents: a single support agent deployment, for example, can connect to each customer's ticketing system using that customer's credentials and apply that customer's specific approval rules. The application authenticates the user, identifies the tenant, and eve loads memory, connection credentials, and approval policy for that user-tenant combination before each call to the model.
For those building B2B products with an embedded agent, this takes off the front-end layer the responsibility of orchestrating which credential and which policy apply to each session open on the customer's screen.
Turn latency dropped from seconds to milliseconds
Vercel published numbers from a hosted stress test comparing the beta (September 2026) with version 1.0 (November 15, 2026), using a load of 100 turns with a mocked model response and no tool calls, isolating the overhead of eve itself and of Workflow.
| Metric | Beta (Sep 2026) | eve 1.0 (Nov 2026) |
|---|---|---|
| Median latency per turn | 3.036s | 414ms |
| p95 | 3.692s | 633ms |
| Estimated cost (50-turn session, 30-day retention) | 5.4¢ | 1.3¢ |
The latency improvement came from joint work with the Workflow team on something Vercel calls spec 9: each execution now has a single orchestrator, so background steps no longer replay the entire workflow. Event and stream writes moved to WebSockets, and redundant history reads were cut. eve also started caching the information needed to route each session's messages and can move idle sessions to a newer deployment between turns; web UIs can pre-warm a session before the first message arrives, which speeds up the first response.
For those measuring perceived performance in a product with embedded AI, this last point matters the most: the time until the first response appears on screen tends to matter more to the user than the total latency of the conversation.
Cost also dropped, and not only on Vercel's hosting
The cost reduction comes from Workflow SDK 5.5, which now saves only what changed between turns instead of copying the entire growing conversation at every step. Checkpoints store references to image files returned by tools, without copying the image bytes each time, and sandboxes built from prepared snapshots skip repeated setup steps.
Since these storage changes are part of the open-source Workflow SDK, self-hosted agents on other workflow backends also store less duplicate session data, not just those using Vercel's managed hosting.
Cross-session memory and observability with OpenTelemetry
eve 1.0 adds native memory for the agent to carry context from one session to another, with configurable scope. Besides the built-in fileMemory (stored in-memory in eve dev and in a private Vercel Blob in production), the list of providers includes claude-mem, Hindsight, Kybernesis Arcana, Memanto, memco, Supermemory, and Upstash AgentKit.
The tracing part was rebuilt on top of OpenTelemetry. Traces now show when an agent is waiting on a person (approval, question, budget prompt, login), log model selection and approval decisions alongside usage and cost, and follow calls down to MCP servers and remote agents. Those hosting on Vercel see all of this in the Agent Runs dashboard, already configured automatically when eve deploy creates a new project.
Who's already using it in production
Vercel cites concrete cases in the announcement: Rillet built a "software factory" on top of eve that turns customer requests into specs, tests the changes, and opens pull requests for human review, and tripled its delivery rate in three months. Clay built monty, an agent that answers employee questions on Slack, and two months after launch it was already used weekly by 30% to 40% of the company.
Custom agents built with eve plus Vercel workflows. I'm biased but I didn't mandate this, we felt it was the best solution for us.
Mitchell Hashimoto, co-founder of HashiCorp
What breaks, and when it's not worth it
Those who already have an eve agent in production need to follow the official upgrade guide: version 1.0 changes how extensions are built, the agent APIs, session events, clients, and saved history. The project follows semantic versioning, so breaking changes only happen in major releases; experimental and internal APIs can change between major versions without prior notice.
It's also worth asking whether the entire harness is necessary. If the use case is a simple chat, a prompt and a response, with no tools, no human approval along the way, and no need to resume execution after a crash, much of eve's complexity (subagents, memory, multi-tenant, durable execution) is overhead that a thinner layer, like plain AI SDK, solves with fewer moving parts.
For those building this kind of interface only occasionally, the simpler path is worth trying first; eve pays off when the product already depends on approval, delegation to subagents, or sessions that need to survive restarts, a scenario where building this engineering by hand would cost much more development time than adopting the ready-made framework.
Getting started
To run locally, you need Node.js 24 or newer:
npx eve@latest init my-agentIn an interactive terminal, this installs the dependencies and opens eve's terminal UI; with access to the AI Gateway through the Vercel team, you can already chat with the agent without creating a project or copying an API key. There are also ready-made templates for chat interfaces in Next.js and TanStack Start, a voice agent, and a Python agent, plus company brain, data analysis, and memory templates, listed in eve's documentation.
Translated from the Brazilian Portuguese original · Read the original
Seedance 2.5 Flash joins Vercel's AI Gateway catalog for cheaper AI video
ByteDance launched a lighter version of Seedance 2.5, and Vercel has already made the model available in the AI Gateway, integrated with AI SDK 7. For those building generative video features, this means less latency and lower cost per clip, without switching providers.





