Skip to content
ISSUE 001·LIVE·13:04 IL
← Journal/2026-11-17·9 min·vercel

Vercel's eve 1.0 is GA: an open-source agent framework with a harness you own

Vercel released eve 1.0, its open-source framework for production agents, after 230-plus releases since the June preview. The framework gives you the agent loop, durable execution, multiplayer sessions, native messaging channels, memory, and OpenTelemetry tracing, with reported median turn latency down from 3.04s to 414ms and workflow costs down about 76%. Apache-2.0, self-hostable, with Vercel as the managed path.

By Harel Asaf·AI Builder·Tel Aviv

Vercel shipped eve 1.0 today, and it is the most complete open-source answer so far to the question every team building agents hits: who owns the loop. eve gives you the harness (the loop that calls models, runs tools, and delegates to subagents) as an Apache-2.0 framework you can read, modify, and self-host, with durable execution that survives crashes, native Slack and other messaging channels, per-user approvals, memory, and OpenTelemetry tracing built in. Vercel hosting is the managed path, not a requirement. The launch numbers are strong on paper: median turn latency down from 3.036s in September to 414ms, and estimated workflow cost for a 50-turn session down about 76%. Those are Vercel's own measurements, with assumptions I will flag below. Everything performance-related in this article is vendor-reported.

What eve 1.0 actually is

eve is a TypeScript framework for building production agents. The core pieces:

  • The harness: the agent loop that calls models, runs tools, and delegates work to subagents. You choose the model, add or replace tools, edit instructions, and decide which actions need approval.
  • Durable execution: the open-source Workflow SDK saves progress as the agent runs, so a crashed agent picks up from its last completed step. Tools can wait for a person, a webhook, or a timer without keeping compute running.
  • Channels: native Slack, Microsoft Teams, Discord, Telegram, GitHub, Linear, Twilio, Web Chat, and iMessage, plus any Chat SDK adapter (WhatsApp, Messenger, Google Chat, email). Each channel is one file, and the same agent works across all of them.
  • Extensions: shareable packages of tools, skills, instructions, connections, channels, schedules, hooks, and subagents, so teams stop copying files between agents.

The repo (vercel/eve) is Apache-2.0, created in June 2026, and currently sits around 5,900 stars. Since the June public preview, Vercel says it shipped 230+ releases and merged thousands of pull requests to reach 1.0.

The 1.0 numbers, and how to read them

Two headline measurements from the launch post:

MeasurementSeptember beta1.0 (November 15)
Median turn latency (hosted stress test)3.036s414ms
p95 turn latency3.692s633ms
Estimated Workflows cost, 50-turn session5.4 cents1.3 cents

The latency test sends messages one at a time to an agent that returns one immediate mock-model response with no tool calls. It measures eve and Workflow overhead, deliberately excluding model generation time. So 414ms is framework overhead per turn, not agent responsiveness as a user experiences it. The cost estimate assumes roughly two model steps per turn and 30-day retention on Vercel Workflows pricing.

The engineering behind the drop: Workflow spec 9 gives each run a single orchestrator so background steps stop replaying the workflow, event and stream writes moved to WebSockets, redundant history reads were cut, and steps now save only the conversation data that changed instead of copying a growing transcript into every step. The storage changes are in the open-source Workflow SDK, so self-hosted agents on other Workflow backends benefit too.

Multiplayer sessions and the channel story

The most opinionated part of eve is how agents meet people. Multiplayer lets several people work with one agent in a shared session: eve tracks who said what and routes approvals and sign-ins to the right person. In the native Slack channel, each thread is one shared session, approvals appear as buttons only the right person can answer, and a live task card shows progress.

Ambient mode is the interesting bet: an agent listens in channels you choose without being mentioned, guided by per-channel goals, and decides when to speak and when to stay quiet. A small decision model screens every message first to limit model costs, and only messages that might be for the agent start a turn. That is the right architecture for cost control, and also the part I would test hardest before pointing it at a busy company channel.

The builder surface

#### Workflow tools as public API

eve's own subagents, sleep, and question tools are built on the same workflow system you use through defineWorkflowTool. A tool can return immediately with a receipt while its task body waits for a subagent's result in a later message, surviving restarts in between. You do not need a separate background-job system.

#### Multi-tenant agents

Dynamic definitions load the agent's capabilities (models, connections, subagents, tools, skills, instructions) from the session before each model call, so one support-agent deployment can connect to each customer's systems with that customer's credentials and approval rules.

#### Memory

Native memory with a scope you choose: built-in file memory (in-memory in dev, a private Vercel Blob store on Vercel), or providers including claude-mem, Hindsight, Kybernesis Arcana, Memanto, memco, Supermemory, and Upstash AgentKit.

#### Tracing with privacy controls

Tracing was rebuilt around OpenTelemetry. Traces show when an agent is waiting for a person (approvals, questions, budget prompts, sign-ins) and record model selections, approval decisions, usage, and cost, following calls into MCP servers and remote agents. You control what gets collected per destination: excluded content stays out of every destination, and private production conversations send only metadata by default.

Coding agents are a first-class template

The 1.0 code extension builds on the default file tools, adds coding instructions and skills that follow your repository's AGENTS.md, and runs the project's checks in a sandbox. The companion git extension adds GitHub and pull request workflows with credentials you configure. The remote self-modification extension lets an agent propose changes to its own source as a draft pull request for human review, which is the right default for that capability. Vercel references benchmark comparisons against other coding harnesses using the same tasks and model; I did not review that benchmark page, so treat the claim as unverified here.

Vercel also quotes production users: Rillet built a software factory on eve and tripled its shipping rate in three months; Clay's data agent reached 30 to 40% weekly company usage in two months; an Evisions engineer reported cutting 25,000 lines of code and halving AI costs after migrating. These are customer stories in a launch post, so read them as marketing with real names attached, and Lindy's agent framework roundup naming eve "Best for durable backend agents" is one roundup's opinion.

The Vercel tie-in, honestly

eve is genuinely open source and self-hostable, but the launch is also a Vercel funnel, and it is worth being clear-eyed about which parts are framework and which are platform:

  • Framework (Apache-2.0, runs anywhere): the harness, channels, extensions, Workflow SDK's durable execution and storage improvements, memory interfaces, tracing.
  • Vercel platform (managed, paid): Vercel Workflows for durable execution at scale, Vercel Sandbox for isolated code execution, Agent Runs dashboards, AI Gateway for model access without your own keys, Fluid compute for workflow steps.

You can start locally with npx eve@latest init my-agent (Node.js 24+) and use existing Vercel CLI credentials for AI Gateway locally. Deploying to Vercel is one command, and that is exactly the point: the friction-free path leads to Vercel's metered services. Self-hosting means you operate the workflow backend and sandboxing yourself.

What I could not verify

All latency and cost numbers come from Vercel's own stress test and pricing estimates, with mock models and stated assumptions; I did not reproduce them. The customer metrics (Rillet, Clay, Evisions) are quotes in the launch post. The coding-harness benchmark comparison is referenced but I did not read the underlying evaluation. The 230+ release count and 5,900 stars are from the announcement and the GitHub repo respectively; the repo is public and active with over 900 open issues, which cuts both ways: real usage, and real bug surface. eve only reached 1.0 today, so production track record at this exact version is zero by definition.

Frequently asked questions

Is eve 1.0 open source?

Yes. The framework and the Workflow SDK are Apache-2.0 on GitHub (vercel/eve). You can self-host, modify, and extend it without Vercel. The managed conveniences (Workflows hosting, Sandbox, Agent Runs, AI Gateway) are Vercel platform services.

Do I have to host eve agents on Vercel?

No. eve dev runs locally, and the Workflow SDK's durable execution and storage improvements are open source, so agents can run on other workflow backends. What you give up off-Vercel is the managed sandbox, one-command deploys, dashboards, and AI Gateway.

What makes eve different from LangChain or the OpenAI Agents SDK?

The durable-execution core and the channel system. Workflow tools that wait on people, webhooks, or timers without holding compute, crash recovery from the last completed step, and native multiplayer messaging channels with per-user approvals are built in rather than assembled. The tradeoff is a younger ecosystem and a framework that reached 1.0 today.

How does eve handle approvals and security?

You decide which actions need approval, and approvals route to the correct person in multiplayer sessions, as Slack buttons only that person can press. For tracing, you choose what is collected per destination; private production conversations send metadata only by default. Agent-generated code runs in an isolated sandbox.

What does multiplayer actually mean in eve?

Several people share one agent session. eve tracks authorship per message, directs approvals and sign-ins to the right individual, and supports per-user or per-team memory. In Slack, one thread equals one shared session with a live task card.

Can eve agents survive crashes?

That is the design center. The Workflow SDK checkpoints progress as the agent runs, approvals and finished steps resume cleanly after a restart, and tools can suspend while waiting for external events without holding compute. One quoted engineer describes stress-testing crash recovery with clean resume, but that is anecdote, not a guarantee.

How much does it cost to run?

Self-hosted, the framework is free and you pay your model and infrastructure bills. On Vercel, Vercel's estimate for a 50-turn session's workflow cost is about 1.3 cents after the 1.0 storage changes, assuming two model steps per turn and 30-day retention; model calls, sandbox time, and AI Gateway usage are separate metered costs.

Sources

Related reading

  • GitHub's Agent Host Protocol: one agent session that follows you across devices

    GitHub published the design of the Agent Host Protocol (AHP), the open standard behind Copilot session continuity across VS Code, the Copilot app, phones, and cloud sandboxes. AHP models agent sessions as URI-identified channels with reducer-based state, so any client sees the same agents, history, and state. Client libraries ship in six languages under MIT, and two independent host implementations already run in production.

  • Self-hosted Claude Managed Agents: Anthropic's agent service now runs on your infrastructure

    Anthropic opened self-hosted Claude Managed Agents in private beta: you host the agent service, its database, and the sandboxes, while model calls go to the Claude API or a provider you choose. What runs where, which credentials exist, what still leaves your network, and what to check before committing.

  • Anthropic's always-on agents guide: the five primitives behind agents that never sit idle

    Anthropic published a guide on building always-on agents: stateful, proactive agents that manage context across sessions. The core idea is that the agent is not actually running all the time - it sleeps between runs, and five architectural primitives decide when it wakes, what it remembers, and what reaches you.

The build log

Follow the work.

New essays and prototypes, in your feed reader. Email subscriptions are paused while I connect a working delivery service. No email is collected here.

Follow via RSS