
Why it mattersA team writing its own agent loop can swap to a maintained one. Vercel says the hosted version runs faster and costs less for long sessions.
Vercel released eve 1.0 on 17 November. It is an open-source framework for building AI agents. The post is signed by nine Vercel engineers, led by Andrew Barba and Shar Dara. The post is on the Vercel blog. A public preview opened in June. Vercel says it has released "230+ releases and merging thousands of pull requests" since then. The framework is now generally available.
Vercel describes eve as the loop that calls models, runs tools and delegates work to subagents. A developer picks the model. A developer writes the tools. A developer sets which actions need approval before they run. The project supports framework integrations for Next.js, Nuxt, SvelteKit and Vite. It supports memory providers claude-mem, Hindsight and Memanto. It supports OpenTelemetry tracing. It has channels for Slack, Microsoft Teams, Discord, GitHub, Linear and Twilio.
Speed and cost figures
Vercel reports its own stress test for the hosted version. Vercel writes that the test sends messages one at a time. Each turn has one immediate mock-model response and no tool calls. The test measures only the time from sending a message until the turn finishes. The number isolates the time eve and Workflow spend handling each turn. It sets the model's own generation time aside. On that test, median turn latency fell from 3.036 seconds to 414 milliseconds.
For long sessions, Vercel gives a cost figure from its own Evo Docs example. The post says the estimated Vercel Workflows cost for a 50-turn session fell from 5.4 cents to 1.3 cents. Vercel describes the drop as "about 76% lower". Both numbers come from Vercel's own tests. Real production cost will depend on model tokens and tool calls. This test does not use either.
Three named customers
The post quotes three customers by name. Rillet says it tripled its release rate in three months after building a software factory on eve. Clay says a data agent named "monty" was used weekly by 30 to 40 percent of the company. Clay said that two months in. Chris Fleissner of Evisions says his team cut 25,000 lines of code. He says his team halved its AI costs and improved runtime by 66 percent. Each figure is each company's own account. None was audited.
Mitchell Hashimoto also appears in the post. He says Vercel built custom agents for his team using eve and Vercel Workflows. He writes that he did not mandate the choice. The team "felt it was the best solution for us".
A team already writing its own agent loop now has a maintained alternative. Vercel publishes its hosted numbers for both. The agent loop is a piece most teams write once and rewrite. Vercel's hosted figures will not match every production workload. A team considering a switch should measure on its own tasks before it moves. The project is open source, so a team can also run it without the hosted runtime.
Source
Introducing eve 1.0, Vercel.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.

