Vercel Workflows officially launched, replacing the entire backend orchestration infrastructure with just two lines of code commands

ME News Report, April 17 (UTC+8), according to Beating Monitoring, the front-end cloud platform Vercel officially released Workflows, integrating durable execution capabilities directly into application code. Developers write "use workflow" at the top of TypeScript functions to mark the workflow entry point, and "use step" within sub-functions to mark each execution step. The framework automatically handles queue scheduling, failure retries, state persistence, and observability, eliminating the need for separate deployment of orchestration services, message queues, or state databases.

The core problem this model addresses is: when moving AI agents or backend tasks from prototype to production, developers often spend a lot of time building orchestration infrastructure instead of improving the product itself. Traditional solutions require breaking logic into queues, workers, state tables, and retry mechanisms. Workflows consolidates all these into application code, integrating orchestration logic with business logic. Vercel states that users only pay for the actual runtime computation of functions, without the overhead of a persistent orchestration service.

Since public testing began in October 2025, Workflows has handled over 100 million runs and 500 million steps, serving more than 1,500 clients, with npm weekly downloads exceeding 200k.

For AI agent scenarios, Workflows offers several key capabilities:

1. Durable Streams: Persistent storage of agent outputs, allowing workflows to continue running after the user closes the browser, and resume from breakpoints upon reconnection, without Redis or custom message systems.
2. Default Encryption: All step inputs, outputs, and stream data are automatically encrypted before leaving the deployment environment, with decryption occurring only within the deployment running the workflow.
3. Suspend and Wake: Wait for external events (like manual approval) via hooks, or pause for days or even months using Sleep, with no compute costs incurred during suspension.
4. High Load Support: Up to 50 MB per step, and a maximum of 2 GB per run, accommodating multi-modal agents passing images and videos.

AI SDK v7 simultaneously launches \WorkflowAgent, deeply integrating durable execution with tool invocation and state management. The Python SDK enters public testing, extending this programming model into the Python ecosystem. The Workflow SDK is open source, supporting self-hosted deployment via the "Worlds" adapter system. The community is developing adapters for MongoDB, Redis, Cloudflare, and others.

Next version, Workflows 5, will introduce native concurrency control (including cross-run locking mechanisms), global deployment infrastructure, and snapshot-based runtime to reduce event replay overhead.
(Source: BlockBeats)