Community Session: Vercel Workflow

Highlights

• Vercel Workflows enables long-running background processes using three primitives: steps, hooks, and sleeps.

• The workflow SDK is open-source and supports portability through adapters for local file systems, PostgreSQL, and upcoming support for Cloudflare and AWS.

• The Durable Agent class within the SDK automates complex agentic loops, including streaming tokens and executing tool calls.

• Performance limits currently track event logs, where high-parallelism tasks exceeding 5,000 steps may experience 15 to 20-second latencies between operations.

• Version 5 of the workflow SDK focuses on improving runtime performance to achieve O(1) suspension and resumption, eliminating replay limitations.

Timeline

1. 00:00:00 - 01:45:00

   Workflow Core Primitives and Agents

   The workflow architecture is designed to model complex background tasks as long-running code. The system manages the complexities of stateful execution, allowing agents to loop through LLM calls and tool usages reliably. This design moves the heavy lifting of orchestration from the developer to the workflow runtime.

   • Workflows rely on three primitives: steps for state, hooks for external events, and sleeps for time-based waiting.
   • Durable Agent classes automate orchestration for repetitive tasks like LLM streaming and tool execution.
   • Background pipelines allow tasks to run across durations ranging from seconds to months.
2. 01:45:01 - 05:27:00

   Portability and Open Source Implementation

   Users can run the same code across different environments by switching the backend via environment variables. The local development experience mocks production behavior by using the local file system for storage and an in-memory queue. This decoupling of the execution logic from the infrastructure layer ensures production workflows remain portable.

   • The workflow SDK is an open-source framework analogous to Docker for long-running code.
   • Existing adapters support local file system inspection, Vercel-hosted encrypted environments, and PostgreSQL persistence.
   • Community and team efforts are expanding support to include Cloudflare and AWS environments.
3. 05:27:01 - 16:09:00

   Agentic Integration and Scaling

   Agents perform well with the SDK because the code relies on idiomatic JavaScript, such as Promise.race patterns for timeouts. While current performance is limited by the size of the event log, upcoming runtime improvements aim to remove the need for 'continue-as-new' workarounds. Developers can currently access existing complex examples, such as flight booking apps and v0's stop-button mechanism, through the public examples repository.

   • AI coding agents natively understand the SDK because it uses standard JavaScript patterns.
   • Workflows can manage concurrency and process termination using primitives like distributed abort controllers.
   • Current event-log based architecture supports high-scale execution but faces latency issues beyond 5,000 steps.
   • Future SDK updates target O(1) suspension and resumption using snapshot-based runtimes.