Vercel shipped eve 1.0 on November 17 — the production-ready, Apache-2.0 open-source agent framework it has been building since June. After 230+ preview releases, the 1.0 marker delivers two things that actually move the needle: median turn latency dropped from 3,036ms to 414ms (a 7.3x improvement), and 50-turn session costs fell from 5.4¢ to 1.3¢ — a 76% reduction. The RC6 build we covered earlier this month was a preview. This is the first version Vercel is calling production-ready.
What eve is, if you missed it
The design premise is simpler than most agent frameworks: an agent is a directory of TypeScript and Markdown files, and you deploy it with one command. No graphs, no DAGs, no “nodes and edges” to wire together. Agents run as standard Node.js processes, so they stay portable. The mental model maps closer to a Next.js app than to a state machine — which is the point for the TypeScript developers Vercel is targeting.
Durability is handled by Vercel’s Workflow SDK. Every turn is checkpointed, so a session can pause for days, survive a deploy, or crash and pick up exactly where it left off. That part comes with a caveat worth naming: the durable execution story requires Vercel infrastructure. You can run eve locally without it, but you lose the crash recovery and long-pause guarantees. For teams already on Vercel, this is a non-issue. For everyone else, it is a tradeoff to evaluate.
How they cut latency by 7.3x
The core problem: each durable checkpoint adds scheduling and persistence latency. In earlier versions, some ordinary turns were paying the full cost of a durable step even when that step did not do meaningful work. The 1.0 fix was architectural — Vercel redesigned core execution mechanics (spec 9), assigning single orchestrators per session and eliminating redundant step replays.
| Metric | Before 1.0 | After 1.0 | Change |
|---|---|---|---|
| Median turn latency | 3,036ms | 414ms | 7.3x faster |
| P95 turn latency | ~2,500ms | ~425ms | 83% reduction |
| 50-turn session cost | 5.4¢ | 1.3¢ | 76% cheaper |
414ms is the threshold that separates “agent feels broken” from “agent feels fast enough to use.” Getting from 3 seconds to under half a second unlocks a category of real-time use cases that were not viable before.
What’s new in 1.0
Code extension and AGENTS.md
The 1.0 code extension gives any eve agent coding capability. It reads AGENTS.md from your workspace root — and falls back to CLAUDE.md if there is not one. Vercel is explicitly endorsing AGENTS.md as the inter-toolchain standard for agent instructions, the same pattern Anthropic uses with CLAUDE.md for Claude Code. The more frameworks converge on this file, the more useful it becomes.
Agent2Agent (A2A) v1.0
Agents can now be published as Agent2Agent (A2A) v1.0 servers using a2aChannel from eve/channels/a2a. Other agents discover them via an Agent Card and delegate durable, streaming tasks over JSON-RPC. eve is one of the first production frameworks to ship native A2A support — a bet on the protocol becoming the standard for agent interoperability.
OpenTelemetry tracing
Tracing is rebuilt around OpenTelemetry, so you can route agent observability to whatever tools you already use. Traces show agent waits — tool approvals, budget prompts, sign-ins — alongside model selections and cost, and they follow calls into MCP servers and remote agents. This is the observability that actually helps when something breaks in production.
Multiplayer sessions
Several people can now share a single agent session. eve tracks who said what and routes approvals to the right person — useful for team debugging and for teams where agents are shared resources rather than personal assistants.
eve vs LangGraph vs CrewAI
Pick eve if you are building TypeScript agents on serverless infrastructure and need production durability. Pick LangGraph if you need fine-grained graph control and time-travel debugging in Python. Pick CrewAI if you want the fastest path to a role-based multi-agent Python setup with enterprise adoption behind it. These frameworks solve meaningfully different problems — the question is which fits your language, infrastructure, and team model.
Getting started
eve is open source on GitHub under Apache-2.0. Local development is free. Hosted deployment runs on Vercel’s standard pricing — eve itself costs nothing, but you are metered on functions, workflows, sandbox time, and model tokens separately. The 1.3¢ per 50-turn session is Vercel Workflows cost only; model token cost is on top.
npx eve@latest init my-agent
cd my-agent
npx eve@latest dev
The full 1.0 announcement has the complete changelog from RC6 and migration notes for teams upgrading from preview builds.

