You can just ship agents
By Steven Van ·
The new Workflow Development Kit adds durable, resumable TypeScript agents to Vercel's AI Cloud, alongside AI Gateway and Sandbox.
Vercel released the Workflow Development Kit, a new way to write durable agents directly in TypeScript, and laid out the AI Cloud platform pieces it says agents need to run in production, in a post today.
The kit lets developers mark boundaries in code that should persist, and define how each step retries or fails. Deployed to Vercel, a workflow saves its progress at each step, can resume after a crash or deploy, and exposes traces of what happened at each step.
Vercel described four pieces of AI Cloud that support this:
- AI Gateway: one key for many models and providers, with automatic failover and visibility into latency and usage
- Fluid compute: scales up while an agent is thinking and to zero while it waits, so billing covers only active CPU time
- Sandbox: runs generated or untrusted code in microVMs that are fast to start
- Workflow tooling: the durability the Workflow Development Kit expresses in code, instead of a separate scheduler or queue
Vercel also said its AI SDK, its TypeScript library for defining agent logic and switching between model providers, is the second most popular JS/TS package for AI programming behind OpenAI's, and that over 3 million developers download it weekly from npm. As examples, it cited internal agents including a support agent that cut ticket volume by a third and a content agent that drafts blog posts from Slack threads.