--- schema: "swft.publication/v1" id: "company-vercel-eve" title: "Vercel eve: the shared infrastructure behind more than 100 production agents" description: "How Vercel uses the open-source eve framework for production agents, durable workflows, sandboxes, permissions, evals, previews, and human approval." summary: "Vercel eve is a shared framework for many kinds of production agents, including software work. It gives teams one file-based convention for resumable work, isolated computers, approved tool access, activity records, repeatable behavior checks, previews, human approval, and reversing a bad release. Vercel says more than 100 internal agents run on it." canonical: "https://swft.io/companies/vercel-eve" author: "SWFT Editorial" author_type: "Organization" published: "2026-09-01" modified: "2026-09-02" kind: "case-study" section: "Companies" tags: ["Vercel eve", "agentic infrastructure", "AI agent framework", "durable agents", "software factory"] evidence_labels: ["INFERENCE", "OBS", "SELF-REPORT"] source_ids: ["vercel-agent-responsibly", "vercel-enterprise-agents", "vercel-eve-factory-template", "vercel-eve-github", "vercel-introducing-eve"] authorship_disclosure: "AI-drafted from the cited public sources and independently checked by a second AI editorial-review agent (Codex) for source fit, claim boundaries, overlap, and reader utility. SWFT Editorial is responsible for corrections." --- # Vercel eve: the shared infrastructure behind more than 100 production agents Vercel standardized file-based agent definitions, resumable runs, sandboxes, permissioned connections, activity records, repeatable behavior checks, previews, and rollback in one platform. > **Authorship:** AI-drafted from the cited public sources and independently checked by a second AI editorial-review agent (Codex) for source fit, claim boundaries, overlap, and reader utility. SWFT Editorial is responsible for corrections. ## Quick answer Vercel eve is a shared framework for many kinds of production agents, including software work. It gives teams one file-based convention for resumable work, isolated computers, approved tool access, activity records, repeatable behavior checks, previews, human approval, and reversing a bad release. Vercel says more than 100 internal agents run on it. Vercel's lowercase **eve** is broader than a coding agent. It is the shared framework and production substrate Vercel says it uses for more than 100 agents across support, sales, engineering, and daily operations. ## What is Vercel eve? [eve](https://vercel.com/blog/introducing-eve) is an open-source framework for defining, running, and operating agents. Vercel says employees had already shipped hundreds of internal agents and apps, each with its own version of scheduling, state, tools, permissions, and monitoring. eve turns those recurring needs into a standard platform. An eve agent is a directory of files. The code, instructions, tools, skills, smaller agents, channels, and schedules live together and can be reviewed in source control. Durable workflows save checkpoints, so a long run can survive a crash or deployment. Generated code runs in an isolated sandbox. OpenTelemetry, an industry standard for recording what software did, exposes each step. Evals, or repeatable checks of agent behavior, can run locally or in continuous integration (CI), the automated build-and-test process for proposed code. This is **agentic infrastructure**: the shared machinery that lets many agents operate without every team rebuilding identity, execution, memory, observability, and release controls. ## How does Vercel use eve internally? Vercel says eve is the framework it builds and runs its own agents on. Its [launch account](https://vercel.com/blog/introducing-eve) names more than 100 agents in production. Examples include d0 for internal questions, a lead agent for sales, Athena for revenue operations, Vertex for support, and coding-related agents such as draft0 and V. These examples matter because the platform serves many business workflows. A software factory is one application of the substrate. Vercel's published outcome figures for named agents are company self-reports with different denominators and acceptance rules, so they should be read individually rather than combined into one productivity claim. ## How do permissions and human approval work? The open framework can pause an action for a person's approval without keeping compute running. Vercel's managed stack adds private-by-default internal apps, employee identity, short-lived task-scoped connections to services, and audit trails. The launch material describes connections to systems such as Slack, GitHub, Snowflake, Salesforce, and Linear. The boundary matters. eve itself is Apache-2.0 open source. Some identity, connectivity, and hosting controls are managed Vercel services and were in beta or private beta when announced. The whole operating stack is not an open-source package. ## What does eve look like as a software factory? Vercel Labs publishes an [eve Software Factory Template](https://github.com/vercel-labs/eve-software-factory-template) called Foreman. A classifier routes work, an analyst researches it, an implementer changes code, and an independent reviewer inspects the pushed branch. Red CI can send work back through the loop. A person reviews, marks the pull request ready, and merges it. Foreman is a useful reference design because every station has its own agent and sandbox, and the reviewer sees the artifact instead of the implementer's private reasoning. Public evidence does not show that Vercel's own engineering organization runs Foreman internally. It is a template built on the same substrate. ## How does Vercel describe safe operation? Vercel's [safety writing](https://vercel.com/blog/agent-responsibly) says green CI is insufficient. An engineer should understand and own agent output. It describes canary releases, which expose a change to a small audience first, plus rollback, runtime verification, executable skills and guardrails, static checks, production-like end-to-end tests, read-only verifier agents, and defect measures. Some of that language describes current capabilities and some describes investment direction. Support for a control does not prove that every internal eve agent uses it. ## What can a product team copy? Standardize the recurring substrate once several agents reach real use. Keep definitions in reviewable files. Give long runs durable state and isolated execution. Broker credentials outside the model. Trace every action. Make evals part of development and CI. Reuse the ordinary preview, approval, release, and rollback path. The [eve repository](https://github.com/vercel/eve) identifies the framework as beta. Its conventions and APIs can still change. Vercel's platform-wide agent deployment statistics also describe all customers on Vercel, not Vercel's internal engineering output. ## How we know - **First-party report (SELF-REPORT)** Vercel describes eve's internal use, production-agent count, conventions, durable execution, sandboxing, connections, observability, evals, and named outcomes. Sources: [Introducing eve](https://vercel.com/blog/introducing-eve); [Vercel for Enterprise Apps and Agents](https://vercel.com/blog/vercel-for-enterprise-apps-and-agents). - **Observed artifact (OBS)** The public eve repository is Apache-2.0 and beta; the Foreman repository is an MIT-licensed reference factory with a human merge gate. Sources: [eve: The Framework for Building Agents](https://github.com/vercel/eve); [eve Software Factory Template](https://github.com/vercel-labs/eve-software-factory-template). - **First-party report (SELF-REPORT)** Vercel states that engineer ownership, runtime checks, release controls, and defect measurement are necessary beyond green CI. Sources: [Agent responsibly](https://vercel.com/blog/agent-responsibly). - **Analysis (INFERENCE)** SWFT treats eve as a factory substrate that standardizes recurring agent operations rather than a single coding product. Sources: [Introducing eve](https://vercel.com/blog/introducing-eve); [eve Software Factory Template](https://github.com/vercel-labs/eve-software-factory-template). ## Sources - **First-party report (SELF-REPORT)** [Agent responsibly](https://vercel.com/blog/agent-responsibly) — Vercel; published 2026-03-30; accessed 2026-09-01. Vercel's stated safety approach, including engineer ownership, release controls, runtime checks, executable guardrails, and defect measurement. - **First-party report (SELF-REPORT)** [Vercel for Enterprise Apps and Agents](https://vercel.com/blog/vercel-for-enterprise-apps-and-agents) — Vercel; published 2026-06-16; accessed 2026-09-01. Vercel's description of its internal agent adoption and managed identity, short-lived access, private-app, and audit controls. - **Observed artifact (OBS)** [eve Software Factory Template](https://github.com/vercel-labs/eve-software-factory-template) — Vercel Labs; accessed 2026-09-01. A public MIT-licensed reference pipeline called Foreman. It is an illustrative template, not evidence that Vercel uses Foreman internally. - **Observed artifact (OBS)** [eve: The Framework for Building Agents](https://github.com/vercel/eve) — Vercel; accessed 2026-09-01. The Apache-2.0 eve repository and beta warning. Public APIs and behavior may change before general availability. - **First-party report (SELF-REPORT)** [Introducing eve](https://vercel.com/blog/introducing-eve) — Vercel; published 2026-06-17; accessed 2026-09-01. Vercel's account of the lowercase eve framework, its internal production use, filesystem convention, durable runs, sandboxes, connections, traces, evals, previews, and named agents. ## Read next - [Cloudflare's AI engineering stack: a control plane for company-wide agents](/companies/cloudflare-ai-engineering-stack) - [Block's AI software factory: Goose, Builderbot, and the Buzz workspace](/companies/block-builderbot-goose-buzz) - [How an AI software factory works](/software-factory-architecture)