Vercel releases eve 1.0: agent turn latency drops from 3.04 s to 0.4–0.6 s
Vercel has released eve 1.0, an open-source framework that handles everything around the model: crash recovery, action approvals and tracing. In a stress test, platform overhead per agent turn fell from 3.04 s to 0.4–0.6 s, not counting the model's own time. By Vercel's estimate, a 50-turn session now costs 1.3 cents instead of 5.4, a 76% drop. If you're building an assistant or a support agent, you no longer have to write this infrastructure from scratch.
AI-processed from Vercel Blog; edited by Hamidun News
Imagine you've built a support assistant agent. In the demo it answers beautifully, but in real use the server crashes mid-conversation, risky actions need human sign-off, and the model bill grows with every message. You have to build all of that around the model yourself.
Vercel has announced that eve, its open-source framework for such production agents, has reached version 1.0 and is generally available. The headline numbers of the release concern the platform, not the model.
"Turn latency" is the time from sending a message to the end of the turn, excluding response generation. In a stress test it fell from 3.04 s in early September to 0.
4–0.6 s, and the estimated cost of a 50-turn session dropped from 5.4 to 1.
3 cents (about 76%). Since the public preview in June, the team has shipped more than 230 releases and merged thousands of pull requests.
What's inside the framework
At the core of eve is the agent "harness": a loop that calls the model, runs tools and hands tasks off to subagents. You choose the model, add or swap tools, edit the instructions, and manually define which actions require human approval. The open-source Workflow SDK saves progress: after a crash, the agent resumes from the last completed step instead of starting over. You can host on Vercel (execution runs in Vercel Workflows, and code produced by the agent runs in isolation in Vercel Sandbox) or on your own infrastructure.
Teams get multi-user sessions. In Slack, each thread is one shared session, and only the person the request is addressed to sees and can click the approval buttons. Ambient mode lets the agent listen to selected channels and decide on its own whether to chime in or stay quiet. To avoid spending money on a large model, a small filter model checks every message first. Besides Slack, eve supports Microsoft Teams, Discord, Telegram, GitHub, Linear, Twilio, web chat and iMessage, and each channel is a single file.
Other new features include memory across sessions (built in or via providers such as Supermemory and Upstash AgentKit), multitenancy (one agent works in each customer's ticketing system with its own credentials and approval rules), extensions for sharing tools and skills, and a code extension for coding agents that respects the AGENTS.md file in the repository. Tracing has moved to OpenTelemetry. For private conversations in production, only metadata is sent by default, without prompts or tool results.
Speed and cost
The test is simple: 100 consecutive turns, each with an instant response from a mock model and no tool calls. It therefore measures only the overhead of eve and Workflow. The median in 1.0 is 414 ms in the results table and 626 ms on the chart next to it. Either figure is a speedup of roughly 5–7x (from 3.036 s). The main credit goes to the new Workflow specification: each run has a single orchestrator, and background steps no longer "replay" the workflow from the beginning. Events are also written over WebSocket, and there are fewer redundant history reads.
The cost story is similar. The estimate for a 50-turn Evo Docs session (about two model steps per turn, 30-day storage) fell from 5.4 to 1.3 cents. Previously, the growing conversation was copied into every step. Now only the changes between turns are saved, and images in checkpoints are stored as references. Tool lists no longer bloat the initial prompt, and markup for the provider's cache has become more precise. An Evisions engineer reported that moving to eve let them remove 25 thousand lines of code, cut AI costs in half and speed up work by 66%.
What to do with this
According to Vercel, Rillet built a "software factory" on eve that turns customer requests into specs, tests and PRs for review, and tripled its shipping speed in three months. At Clay, the monty agent answers employee questions in Slack. A little over two months in, 30–40% of the company used it every week. This is the vendor's own data, and the latency test is synthetic: real turns with a live model and tools will be slower.
The practical takeaway: if you're building an internal assistant, a support agent or a coding agent and don't want to write your own infrastructure for retries, approvals and tracing, eve is worth trying on a single use case. The benefit is greatest on the Vercel stack, but the savings on session storage also apply on other Workflow backends if you self-host.
- Nov 18, 2026JetBrains' Mellum Runs Locally in an IDE: 5 Out of 10 on a MacBook M3 Pro
- Nov 18, 2026OpenAI Releases Legal AI: 95% Fewer Fabricated Facts Than Claude Opus 5.5
- Nov 18, 2026Claude SEO Plugin Runs Up to 17 Agents and Audits a Site with One Command
- Nov 18, 2026Cloudflare open-sources Kitesurf: AI agents get by on 58 MiB of memory instead of Chromium's 399
Key Questions Answered
What's inside the framework?
At the core of eve is the agent "harness": a loop that calls the model, runs tools and hands tasks off to subagents. You choose the model, add or swap tools, edit the instructions, and manually define which actions require human approval. The open-source Workflow SDK saves progress: after a crash, the agent resumes from the last completed step instead of starting over.
What are the key details on Speed and cost?
The test is simple: 100 consecutive turns, each with an instant response from a mock model and no tool calls. It therefore measures only the overhead of eve and Workflow. The median in 1.0 is 414 ms in the results table and 626 ms on the chart next to it. Either figure is a speedup of roughly 5–7x (from 3.036 s). The main credit goes to the new Workflow specification: each run has a single orchestrator, and background steps no longer "replay" the workflow from the beginning.
What to do with this?
According to Vercel, Rillet built a "software factory" on eve that turns customer requests into specs, tests and PRs for review, and tripled its shipping speed in three months. At Clay, the monty agent answers employee questions in Slack. A little over two months in, 30–40% of the company used it every week. This is the vendor's own data, and the latency test is synthetic: real turns with a live model and tools will be slower.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.