Vercel moved eve out of preview and into general availability with version 1.0, the company announced on its blog. Eve is an open-source framework that runs an agent’s model-and-tools loop, hands work to subagents, and wraps it in memory, approval controls and monitoring. Vercel says the preview, which opened in June, produced more than 230 releases and thousands of merged pull requests before the 1.0 tag.

The core idea is that the team owns the loop. You pick the model, swap or add tools, edit instructions, and mark which actions need a human sign-off. Progress is saved by the open-source Workflow SDK, so a crashed agent resumes from its last finished step. Hosting on Vercel adds Vercel Workflows for durable execution and Vercel Sandbox, an isolated environment for code the agent writes.

Speed is the headline number. Vercel’s own stress run on its Workflows platform put median turn latency at 3.036 seconds on September 1, the beta baseline. On November 15, the 1.0.0 build handled the same 100-turn workload in 414 milliseconds, with p95 falling from 3.692 seconds to 633 milliseconds. That is the 7.3x speedup Vercel cites. The test uses one instant mock-model reply and no tool calls, so it measures framework overhead and says nothing about a real agent waiting on a real model. One chart caption in the post also gives 626 milliseconds for the same measurement, a mismatch the company has not explained.

Cost gets a separate claim. For a 50-turn session on Evo Docs, Vercel estimates workflow spend drops from 5.4 cents to 1.3 cents, about 76 percent lower, assuming about two model calls in each turn plus a 30-day retention window. The savings come from Workflow SDK 5.5, which saves just the parts of a conversation that differ from the prior turn, plus checkpoints that reference image files instead of copying them. Because those changes live in the open-source SDK, people hosting elsewhere get the same reduction in duplicate session data, though the dollar figure uses Vercel pricing.

Customer evidence comes from Vercel itself. Rillet built a software factory on eve that converts customer requests into specs, tests the code, and opens pull requests that people review. Rillet says it now ships at triple the rate after three months. Clay made monty, an agent that answers staff questions in Slack, and 30 to 40 percent of the company used it weekly just over two months in. An engineer at Evisions, Chris Fleissner, described cutting 25,000 lines of code, halving AI costs and improving runtime by 66 percent during an earlier migration. Mitchell Hashimoto, quoted on the page, said his team builds custom agents with eve and Vercel Workflows and was not told to.

Several features target teams rather than solo builders. Multiplayer sessions let several people share one agent while eve sends each approval or login prompt to whoever should answer it. Native channels cover Slack, Microsoft Teams, Discord, Telegram, GitHub, Linear, Twilio, Web Chat and iMessage, and a Chat SDK channel extends that to WhatsApp, Messenger, Google Chat and email. Ambient mode lets an agent sit in a Slack channel unprompted, with a small model screening each message to keep costs down. Multi-tenant support lets one deployment use each customer’s credentials, memory and approval rules.

Memory is now built in, scoped as you choose. Options include a file store, backed by Vercel Blob in production, or providers such as Supermemory, Hindsight, Memanto, memco, claude-mem, Kybernesis Arcana and Upstash AgentKit. Tracing was rebuilt on OpenTelemetry, so spans flow to whatever backend a team already runs, including Raindrop. Excluded content stays out of every destination, and private production conversations send only metadata by default.

Framework bindings cover Next.js, Nuxt, SvelteKit, TanStack Start and Vite. Setup needs Node.js 24 or newer and one command, npx eve@latest init. A coding extension ships in the package, and Vercel points to benchmark results against other coding harnesses on the same tasks and model. The post links those results rather than stating them, so they need independent checking. Eve follows semantic versioning, so stable APIs change only in major releases, while experimental ones can shift between them.

For teams already piloting agents on a homegrown loop, the useful test is narrow: replay your own multi-tool workload on eve and compare turn time and per-session cost against the mock-model numbers before moving anything.

Vercel, “Introducing eve 1.0,” vercel.com blog, November 2026.