deniz.in

Markets

Weather

Loading weather

· via Vercel blog

Vercel's open-source agent framework eve reaches 1.0 and general availability

Vercel has released eve 1.0, moving its open-source agent framework to general availability with much lower turn latency, OpenTelemetry tracing, native memory and multi-tenant support.

Vercel's open-source agent framework eve reaches 1.0 and general availability

Vercel has released eve 1.0, its open-source framework for building production agents, announcing general availability on November 17. According to the Vercel blog, the milestone follows a public preview that opened in June, during which the team shipped more than 230 releases and merged thousands of pull requests while working with early adopters.

A harness developers own

eve is organised around a harness: the loop that calls models, runs tools and delegates work to subagents. Developers choose the model, add or replace tools, edit instructions and decide which actions need approval. The open-source Workflow SDK saves progress as the agent runs, so it can resume from its last completed step after a crash, and multiplayer sessions and agent-to-agent calls are built in. Agents can also be deployed to Vercel, where Vercel Workflows handles durable execution and Vercel Sandbox runs agent-generated code in isolation. Security controls determine which actions require approval and what data is collected, and Vercel says a rigorous evaluation suite now measures how easily coding agents can build with the framework.

Evidence from production

Vercel points to businesses already running eve in production. Rillet built a software factory that turns customer requests into specs, tests the changes and opens pull requests for human review, tripling its shipping rate in three months. Clay built monty, a data agent that answers employee questions in Slack; a little over two months in, 30–40% of the company used it weekly. Lindy's roundup of agent frameworks rated eve best for durable backend agents, and Vercel quotes Mitchell Hashimoto saying his team picked eve with Vercel Workflows as the best solution for them — a choice, he noted, not a mandate.

Channels, sessions and UIs

In Slack, each thread is one shared session: approvals appear as buttons only the right person can answer, a live task card tracks progress and timing, and follow-ups work without mentioning the agent again. Ambient mode lets an agent listen in chosen channels without being mentioned; guided by per-channel goals it decides when to speak up, and a small decision model screens every message first so only plausible ones start a turn, containing model costs. Slack sits alongside native channels for Microsoft Teams, Discord, Telegram, GitHub, Linear, Twilio, Web Chat and iMessage, with a Chat SDK adapter adding WhatsApp, Messenger, Google Chat and email. Each channel is one file, and the same agent can serve them all. On the interface side, native integrations for Next.js, Nuxt, SvelteKit, TanStack Start and Vite build the agent alongside an app, client libraries expose tool calls, approval requests and questions consistently across frameworks, and eve can scaffold a chat app or slot its React, Vue and Svelte bindings into an existing UI.

Multi-tenancy and memory

A single support-agent deployment can connect to each customer's ticketing system using that customer's credentials and approval rules. The host application authenticates users and identifies tenants, after which memory, connection credentials and approval policy can be set per tenant and per user. Dynamic definitions load the agent's capabilities — models, connections, subagents, tools, skills and instructions — from session and external data before each model call. Version 1.0 also adds native memory for carrying context between sessions: built-in file memory keeps data local in eve dev and in a private Vercel Blob store on Vercel, and third-party providers include Supermemory, Upstash AgentKit, claude-mem and Hindsight.

Observability and speed

Tracing has been rebuilt around OpenTelemetry, so agent runs flow into the tracing tools teams already use. Traces show when an agent is waiting on a person — approvals, questions, budget prompts, sign-ins — along with model selections, approval decisions, usage and cost, and they follow calls into MCP servers and remote agents. Collection stays under the developer's control: content excluded from collection remains out of every destination, spans can be filtered, inputs and outputs redacted per destination, and private production conversations send only metadata by default. On Vercel, Agent Runs uses the same traces to show sessions, model calls, tool execution, subagent activity and approval decisions in the dashboard.

Latency also fell sharply. Vercel's hosted stress test on Vercel Workflows measured a median of 3.036 seconds per turn in early September; on 1.0, the same 100-turn workload took a median of 414ms. A chart in the same post cites a slightly different comparison — 3.04s down to 626ms, labelled 4.8× faster — so the exact figure depends on the configuration being measured. Both tests use a mock model that replies immediately with no tool calls, isolating framework and workflow overhead from model generation time. The gains come from a single-orchestrator workflow spec that stops background steps replaying the run, WebSocket-based event and stream writes, fewer redundant history reads, cached session routing, migration of idle sessions between deployments, and UI prewarming of sessions before the first message.

Why it matters

The agent-framework field is crowded, and few projects commit to a stable 1.0. eve arriving at general availability — open-source, with durable execution that resumes after crashes, tenant isolation, approval controls and OpenTelemetry-grade observability — targets the gap between agent demos and systems businesses actually depend on. The adoption evidence Vercel cites, from Rillet's shipping rate to Clay's internal usage, is the kind of proof engineering teams want before adopting infrastructure. And the pattern on show — an open harness developers own, paired with paid managed execution on Vercel — is becoming the default shape of the developer-platform business.

  • #ai-agents
  • #open-source
  • #vercel
  • #developer-tools
  • #opentelemetry

Related posts