A New Programming Model for Durable Execution

Vercel Workflows has reached GA as the easiest way to write long-running, durable, reliable, and observable agents and backends. Its starting point is the observation that prototypes run easily locally, but complexity spikes the moment you have to handle failures, restarts, and real traffic.

Since the beta began in October 2025, it has already processed more than 100 million runs and more than 500 million steps, been used by more than 1,500 customers, and recorded more than 200,000 weekly npm downloads. Now, as it makes Workflows generally available for production use, it pushes a model in which the app code itself acts as the orchestrator.

The core architecture rests on three pieces.

• Event log: a single source of truth for execution history that records every step input, output, stream chunk, sleep, hook, and error
• Fluid compute: each step runs as an independent function invocation, with the library handling queue dequeue, state loading, encryption, execution, and handoff
• Vercel Queues: automatically enqueues the next step, and works across Vercel, self-hosted Postgres, and local in-memory environments

In TypeScript, you declare a workflow with "use workflow" and separate units of work with "use step". On the surface it looks like ordinary function calls, but under the hood each step gets automatic retry, state persistence, observability, and durable continuation.

Features for agents are also highlighted. Durable agents offer deep integration with the AI SDK, supporting state retention, tool calling, handling external events, and resuming after interruption, while Durable streams let a stream keep going even after the browser closes, via getWritable(). With WorkflowChatTransport and x-workflow-run-id, a client that disconnects and reconnects can recover starting from the last event.

On the security side, step input/output/stream chunks are encrypted before leaving deployment boundaries, and viewing them in the dashboard or CLI requires explicit decryption along with an audit trail. It also offers Hooks to wait for external triggers, and Sleep to pause for minutes, days, or months, fitting flows such as human-in-the-loop approvals or email drip campaigns.

Capacity is also designed with agent use cases in mind. It supports up to 50 MB per step and 2 GB per run, targeting multimodal workloads that exchange images, video, and large contexts. The CLI lets you inspect a run with npx workflow inspect runs <wrun_id>, reusing your existing Vercel authentication so you can debug production state right from the terminal.

The deployment model extends through an adapter system called Worlds. Managed has Vercel handle everything, Self-hosted provides a Postgres-based reference implementation along with a Local World, and Embedded lets the community attach additional worlds such as MongoDB, Redis, Turso, Jazz Cloud, and Cloudflare. Customer examples like Mux and Durable show it already being applied to complex production workloads such as video/AI pipelines and website generation.