Since the June public preview, eve has shipped 230+ releases and merged thousands of pull requests. Version 1.0 is generally available, with faster turn handling, broader integrations, rebuilt tracing, native memory, and security controls that decide which actions require approval and what data gets collected. The evaluation suite used to measure the coding-agent experience has also been tightened.
The harness and where it runs
eve's harness is the loop that calls models, runs tools, and hands work to subagents. Every part is yours to change: the model, the tool set, the instructions, and the approval policy. The open-source Workflow SDK persists progress during a run, so an agent resumes from its last completed step after a crash. Multiplayer sessions and agent-to-agent calls ship in the box.
Deploying to Vercel uses Vercel Workflows for durable execution and Vercel Sandbox to isolate code the agent generates.
Production deployments cover software factories that turn requests into tested pull requests, company brains that answer from business data, and customer-facing agents. Rillet's software factory takes customer requests, converts them to specs, tests changes, and opens PRs for human review; the company tripled its shipping rate in three months. Clay's data agent monty answers employee questions in Slack and reached 30–40% weekly usage across the company just over two months after launch. Lindy's framework roundup describes eve as "Best for durable backend agents".
“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↗(opens in new tab)
Shared sessions and channels
Multiplayer sessions let several people work with one agent in the same session. eve tracks authorship, routes approvals and sign-ins to the correct person, and supports memory scoped per user or per team.
In the native Slack channel, each thread is a single shared session: approvals appear as buttons only the right person can press, and a live task card reports progress and timing. Participants can follow up without re-mentioning the agent, cancel the current turn, or reset the session from Slack. Ambient mode goes further, letting an agent listen in chosen channels without a mention and decide from per-channel goals when to speak up, help, or stay quiet. A small decision model screens every message first so only likely candidates start a turn, which limits model costs.
Slack joins Microsoft Teams, Discord, Telegram, GitHub, Linear, Twilio, Web Chat, and iMessage as native channels. The Chat SDK channel covers any Chat SDK platform adapter, including WhatsApp, Messenger, Google Chat, and email. Each channel is a single file, and one agent can serve all of them.
Frontend integration and per-tenant configuration
Native integrations build the agent alongside the app and wire up its routes for Next.js, Nuxt, SvelteKit, TanStack Start, and Vite. On Vercel the agent and site can deploy as one project or separately. Client libraries expose tool calls, approval requests, and questions the same way across frameworks, so session data can drive custom UIs — streamed chat replies, an embedded agent, or the agent's tool calls rendered in your own components. Run eve add channel/web to scaffold a chat app with AI Elements, or add the React, Vue, or Svelte bindings to an existing UI.
Multi-tenancy starts with application-level authentication: your app authenticates the user and resolves their tenant, after which memory, connection credentials, and approval policy can be set per tenant and user. One support-agent deployment can therefore reach each customer's ticketing system under that customer's credentials and rules. Before every model call in the main loop, dynamic definitions load capabilities from session and external data; in 1.0 they cover models, connections, and subagents in addition to tools, skills, and instructions.
Memory across sessions
Native memory carries context between sessions at a scope you choose. Built-in fileMemory holds data in memory under eve dev and in a configured private Vercel Blob store on Vercel. Providers include claude-mem, Hindsight, Kybernesis Arcana, Memanto, memco, Supermemory, and Upstash AgentKit, or you can build your own.
OpenTelemetry tracing
Tracing is rebuilt around OpenTelemetry. eve dev records traces locally by default; self-hosted deployments can export to any OpenTelemetry backend, including one you run yourself. Raindrop is among the available integrations — add one with eve add instrumentation/ or write your own against the instrumentation APIs.
Traces mark when an agent is waiting on a person — tool approvals, questions, budget prompts, sign-ins — and record model selections and approval decisions next to usage and cost. Calls into MCP servers and remote agents are followed as well. Collection is configurable: excluded content never reaches any destination, spans can be filtered and inputs or outputs redacted per destination, and private production conversations send only metadata by default, without prompts or tool results.
On Vercel, Agent Runs renders the same traces in the dashboard, covering sessions, model calls, tool execution, subagent activity, and approvals. When eve deploy creates a new Vercel project it configures trace sampling automatically.
Latency
A hosted stress test on Vercel Workflows in early September measured a median of 3.036s per turn; the same 100-turn workload now takes 414ms. The test sends messages one at a time with a single immediate mock-model response and no tool calls — it times from message send to turn completion, capturing eve and Workflow overhead but not model generation time.
Beta3.04s
1.0414ms
Hosted stress run | Median turn latency | p95 |
|---|---|---|
3.036s | 3.692s | |
414ms | 633ms |
Much of the gain comes from Workflow spec 9, developed with the Workflow team: each run gets a single orchestrator, so background steps no longer replay the workflow, and event and stream writes moved to WebSockets with fewer redundant history reads. Agents inherit this through Workflow SDK. eve additionally caches the routing information it needs per session and can move idle sessions onto a newer deployment between turns. Web UIs can prewarm a session before the first message so the first reply arrives sooner.
Storage and model costs
On a 50-turn Evo Docs session, estimated Vercel Workflows cost falls from 5.4¢ to 1.3¢ — roughly 76% lower — assuming about two model steps per turn and 30-day retention. eve now runs Workflow SDK 5.5 and persists only the conversation data that changed between turns. The savings come from three changes:
- Workflow steps save deltas and reference stored data instead of copying a growing conversation into each step; steps that don't need history no longer receive it.
- Checkpoints store references to tool-returned image files rather than copying image bytes.
- Sandboxes created from prepared snapshots skip repeated setup; a stopped sandbox keeps its files for reuse, and cleanup begins when the session ends.
These storage changes live in the open-source Workflow SDK, so agents self-hosted on other Workflow backends also store less duplicate session data. On Vercel, steps run on Fluid compute and need extra storage writes for inputs over the payload limit, so sending less data keeps turn times steady as sessions grow; images and files may still trigger those writes.
To reduce model spend, tool and skill listings can be withheld from the initial prompt until needed, and connection tools run through eve__search and eve__tool so a search no longer expands the model's tool list each time. Prompt regions are also marked for provider caching differently, so more content is reused between calls. During an earlier migration, an Evisions engineer reported the following:
“Transitioning to Eve delivered immediate, high impact results. To name a few, we were able to cut out 25k lines of code, cut our AI costs in half, and improved runtime by 66%.”
Chris Fleissner
Packaging capabilities as extensions
Extensions, added since the preview, let teams share tools, skills, instructions, connections, channels, schedules, hooks, and subagents without copying files between agents. Use an extension from the extensions directory or create one with npx eve@latest extension init . Examples include agent-browser for browser automation, Mux Video for video assets and clips, and Link for wallet and purchase tools. Workspace extensions are compiled from source and reloaded during development, so a shared capability can be updated in one package and tried immediately across agents.
Coding work inside the sandbox
The 1.0 code extension layers onto eve's default file tools, contributing coding instructions and skills that respect the repository's AGENTS.md. Since it ships inside the eve package, no separate install is needed. Hand an agent a repository in its sandbox and it can explore the code, modify files, and execute the project's checks. Because it uses the same harness and extension system as every other agent, tools remain replaceable or removable, skills can be overridden, and additional instructions can be supplied.
Attach the extension to an agent as follows:
terminal.sh
eve add extension/code
Benchmark comparisons against other coding harnesses, run on the same tasks and model, are published alongside the extension.
A companion module, eve/extensions/git, supplies GitHub and pull request workflows driven by credentials and authorization you set up. On Vercel, an optional remote self-modification extension goes further: the agent proposes edits to its own source as a draft pull request, which you review, merge, and deploy.
Workflows behind the built-in tools
eve's subagents, sleep, and question tools are themselves built on the public workflow system exposed through defineWorkflowTool. A tool defined this way can return a result directly, kick off a background task, or act as a resumable task that keeps accepting messages. Waiting on a person, a webhook, or a timer requires no running compute, and runs persist across restarts, removing the need for an external background-job system to coordinate the work.
Given a reviewer subagent, the tool below hands back a receipt right away while the task body waits for the reviewer's output, which reaches the model in a later message:
agent/tools/review.ts
import { defineWorkflowTool } from "eve/tools";
import { z } from "zod";
export default defineWorkflowTool({
description: "Ask the reviewer to check these notes.",
inputSchema: z.object({ notes: z.string() }),
async task({ notes }, ctx) {
"use workflow";
const response = await ctx.agent("reviewer").send(notes, {
signal: ctx.abortSignal,
});
return await response.result();
},
});
Moving existing agents to 1.0
For agents already on eve, the 1.0 upgrade guide covers the changes affecting extension builds, agent APIs, session events, clients, and saved history, and should be followed before the upgrade is deployed. eve applies semantic versioning, so stable APIs break only in major releases; experimental and internal APIs, experimental fields included, may change between majors.
Existing projects can be upgraded with a coding agent using this prompt (the same instructions apply if you invoke the agent yourself):
Read https://eve.dev/docs/guides/upgrade and inspect this project's eve version and configuration. Upgrade it to eve 1.0, following the guide's order and the sections that apply. Update extensions, agent code, event consumers, clients, and saved history where needed. Run the project's typecheck, eve build, and relevant tests. Summarize the changes and any remaining manual migration or deployment steps.
Three routes to a first agent
You can build locally in the terminal, create and deploy from the browser, or begin with a template.
Terminal
With Node.js 24 or newer, the CLI creates a new agent:
terminal.sh
npx eve@latest init my-agent
Run in an interactive terminal, the command installs dependencies and launches the eve terminal UI. When your team has AI Gateway access, chatting can start without creating a Vercel project or copying an API key, since eve can call models locally through AI Gateway using existing Vercel CLI credentials; /login signs in to other connections. Inside eve dev, the agent can be asked to add a tool, revise its instructions, or write a skill — and with self-modification it edits its own source and reloads it for the next turn.
Browser
The agent builder in the Vercel dashboard scaffolds an agent, creates its Git repo, and deploys it. Choose Add New → Agent, select a model, and optionally add web chat or Slack.
Templates
Ready-made eve templates cover common use cases:
Use case | Templates |
|---|---|
Coding | Coding agent · Software factory · GitHub maintainer · Incident response |
Knowledge and data | Company brain · Data analyst · Shared files with Vercel Drives |
Apps and interfaces | Next.js chat · TanStack Start chat · Voice agent · Agent with Python |
Messaging | Slack · Discord · Personal agent · BELL Stack |
Memory | |
Content and marketing | Contentful · Sanity copilot · Content agent · Marketing team |

Mitchell Hashimoto↗(opens in new tab)
Chris Fleissner

