Recently, I saw that Vercel launched something quite interesting called Workflows, and honestly, it addresses many pain points for backend developers. In simple terms, it replaces complex backend infrastructure with just two lines of code. The core idea is to mark use workflow at the top of a TypeScript function, then mark use step inside sub-functions. The framework automatically handles the rest—queue scheduling, retry on failure, state persistence, and other tedious tasks.

I think the value of this product lies in changing a pain point in the development process. Previously, to move AI agents or backend tasks from prototype to production, developers had to spend a lot of time on infrastructure—things like message queues, state tables, retry mechanisms—all scattered across different places. Now, Vercel integrates all these into the application code, mixing logic and business code, which should greatly improve the development experience.

From the data, since its public testing began last October, Workflows has handled over 100 million executions and 500 million steps, with more than 1,500 customers using it, and npm weekly downloads exceeding 200k. This growth rate clearly indicates demand. Plus, the pricing model is pretty good—charged based on actual execution time, without fixed costs for continuous orchestration services.

For AI agent scenarios, Vercel emphasizes several capabilities. Durable Streams allow the agent’s output to be persistently saved, so it can keep running even if the browser is closed, and reconnect from where it left off. Default encryption is also a selling point— all input, output, and stream data are automatically encrypted before leaving the deployment environment, with decryption only happening inside the execution environment. It also supports pause and resume, for example, waiting for manual approval or sleeping for days or months, with zero compute cost during the pause. The maximum size per step is 50MB, and the total execution can be up to 2GB, which is enough for multimodal agents transmitting images and videos.

The newly released AI SDK v7 integrates WorkflowAgent, deeply combining persistent execution, tool invocation, and state management. The Python SDK is also entering public testing, indicating Vercel’s intention to push this programming model to more developers. The open-source Workflow SDK supports self-hosted deployment, and the community is already developing adapters for MongoDB, Redis, Cloudflare, and others, showing good extensibility.

The next version, Workflows 5, plans to add native concurrency control and global deployment infrastructure, as well as use snapshot runtimes to reduce the cost of reprocessing events. Overall, Vercel’s move truly addresses developers’ pain points by replacing traditional complex backend architectures with a simplified programming model. For teams looking to rapidly iterate on AI agents or asynchronous tasks, this solution is worth trying.