workflow

Critical: Always use correct workflow documentation

Your knowledge of workflow is outdated.

The workflow documentation outlined below matches the installed version of the Workflow SDK. Follow these instructions before starting on any workflow-related tasks:

Search the bundled documentation in node_modules/workflow/docs/:

1. Find docs: glob "node_modules/workflow/docs/**/*.mdx"
2. Search content: grep "your query" node_modules/workflow/docs/

Documentation structure in node_modules/workflow/docs/:

• getting-started/ - Framework setup (next.mdx, express.mdx, hono.mdx, etc.)
• foundations/ - Core concepts (workflows-and-steps.mdx, hooks.mdx, streaming.mdx, etc.)
• api-reference/workflow/ - API docs (sleep.mdx, create-hook.mdx, fatal-error.mdx, etc.)
• api-reference/workflow-api/ - Client API (start.mdx, get-run.mdx, resume-hook.mdx, etc.)
• api-reference/workflow-runtime/ - Runtime API (get-world.mdx) and world/ World SDK (storage.mdx, streams.mdx, queue.mdx)
• api-reference/workflow-observability/ - Hydration and name parsing utilities (hydrate-resource-io.mdx, parse-workflow-name.mdx, etc.)
• ai/: AI SDK integration docs
• errors/ - Error code documentation
• worlds/ - Per-World behavior and limits (vercel.mdx, local.mdx, postgres.mdx). Other pages link these as /worlds/<name>.

Related packages also include bundled docs:

• @ai-sdk/workflow: node_modules/ai/docs/ - WorkflowAgent and AI SDK integration
• @workflow/ai: node_modules/@workflow/ai/docs/ - deprecated DurableAgent APIs for existing applications
• @workflow/core: node_modules/@workflow/core/docs/ - Core runtime (foundations, how-it-works)
• @workflow/next: node_modules/@workflow/next/docs/ - Next.js integration

When in doubt, update to the latest version of the Workflow SDK.

Official resources

• Website: https://workflow-sdk.dev
• GitHub: https://github.com/vercel/workflow

Quick reference

Directives:

"use workflow";  // First line - makes async function durable
"use step";      // First line - makes function a cached, retryable unit

Essential imports:

// Workflow primitives
import { sleep, fetch, createHook, createWebhook, getWritable } from "workflow";
import { FatalError, RetryableError } from "workflow";
import { getWorkflowMetadata, getStepMetadata } from "workflow";

// API operations
import { start, getRun, resumeHook, resumeWebhook } from "workflow/api";

// Observability & data hydration
import { hydrateResourceIO, observabilityRevivers, parseStepName, parseWorkflowName } from "workflow/observability";

// Framework integrations
import { withWorkflow } from "workflow/next";
import { workflow } from "workflow/vite";
import { workflow } from "workflow/astro";
// Or use modules: ["workflow/nitro"] for Nitro/Nuxt

// AI agent (Workflow 5)
import { WorkflowAgent, type ModelCallStreamPart } from "@ai-sdk/workflow";

Prefer step functions to avoid sandbox errors

"use workflow" functions run in a sandboxed VM. "use step" functions have full Node.js access. Put your logic in steps and use the workflow function purely for orchestration.

// Steps have full Node.js and npm access
async function fetchUserData(userId: string) {
  "use step";
  const response = await fetch(`https://api.example.com/users/${userId}`);
  return response.json();
}

async function processWithAI(data: any) {
  "use step";
  // AI SDK works in steps without workarounds
  return await generateText({
    model: "spacexai/grok-4.6",
    prompt: `Process: ${JSON.stringify(data)}`,
  });
}

// Workflow orchestrates steps - no sandbox issues
export async function dataProcessingWorkflow(userId: string) {
  "use workflow";
  const data = await fetchUserData(userId);
  const processed = await processWithAI(data);
  return { success: true, processed };
}

Benefits: Steps have automatic retry, results are persisted for replay, and no sandbox restrictions.

Workflow sandbox limitations

When you need logic directly in a workflow function (not in a step), these restrictions apply:

Example - Using fetch in workflow context:

import { fetch } from "workflow";

export async function myWorkflow() {
  "use workflow";
  globalThis.fetch = fetch;  // Required for AI SDK and HTTP libraries
  // Now generateText() and other libraries work
}

Note: Plain "provider/model" strings use Vercel AI Gateway. Do not construct a direct provider instance unless the user explicitly needs a provider-only feature.

WorkflowAgent: AI agents in Workflow 5

Use AI SDK's WorkflowAgent for durable agents on Workflow 5. It replaces the deprecated DurableAgent API from @workflow/ai and checkpoints model calls and step-backed tools.

import { WorkflowAgent, type ModelCallStreamPart } from "@ai-sdk/workflow";
import { isStepCount, tool } from "ai";
import { getWritable } from "workflow";
import { z } from "zod";

async function lookupData({ query }: { query: string }) {
  "use step";
  // Step functions have full Node.js access
  return `Results for "${query}"`;
}

export async function myAgentWorkflow(userMessage: string) {
  "use workflow";

  const agent = new WorkflowAgent({
    model: "spacexai/grok-4.6",
    instructions: "You are a helpful assistant.",
    tools: {
      lookupData: tool({
        description: "Search for information",
        inputSchema: z.object({ query: z.string() }),
        execute: lookupData,
      }),
    },
  });

  const result = await agent.stream({
    messages: [{ role: "user", content: userMessage }],
    writable: getWritable<ModelCallStreamPart>(),
    stopWhen: isStepCount(10),
  });

  return result.messages;
}

Key points:

• A plain "provider/model" string routes through Vercel AI Gateway; spacexai/grok-4.6 is the default model in Workflow examples
• getWritable<ModelCallStreamPart>() streams durable model-call output; convert it with createModelCallToUIChunkTransform() in an HTTP route
• Tool execute functions that need Node.js/npm access should use "use step"
• Tool execute functions that use workflow primitives (sleep(), createHook()) should NOT use "use step" because they run at the workflow level
• stopWhen limits the number of model calls; the default is to stop when the model stops calling tools
• Multi-turn: pass result.messages plus new user messages to subsequent agent.stream() calls

For more details, check the WorkflowAgent docs in the installed AI SDK package or at https://ai-sdk.dev/v7/docs/agents/workflow-agent.

Starting workflows & child workflows

Use start() to launch workflows from API routes. In Workflow 5, start() can also be called directly from a workflow function to spawn a child run; it is step-backed and records a deterministic boundary in the parent's event log.

import { start } from "workflow/api";

// From an API route; works directly
export async function POST() {
  const run = await start(myWorkflow, [arg1, arg2]);
  return Response.json({ runId: run.runId });
}

// No-args workflow
const run = await start(noArgWorkflow);

Starting child workflows from inside a Workflow 5 workflow:

import { start } from "workflow/api";

export async function parentWorkflow() {
  "use workflow";
  const childRun = await start(childWorkflow, ["some data"]);
  await sleep("1h");
  return { childRunId: childRun.runId };
}

start() returns after creating the child run and doesn't wait for it to complete. Use childRun.returnValue only when the parent should wait for the child; each Run property access or method call inside a workflow is a step.

Run size & concurrency: know when to split

Three things to size, and all three are capped. Do NOT treat any number you remember as authoritative — the current values are published under Workflow run limits, which is the only source to quote.

Events per run. A run's event log is capped, and the run fails with MAX_EVENTS_EXCEEDED past the ceiling. Events are not steps: a step that succeeds on the first try records three (step_created, step_started, step_completed), a retry records one or two more, and hooks, sleeps, and webhooks each record their own. Split into child workflows well before the ceiling — the pricing page recommends that past a few thousand events, because replay slows down long before the run fails.

Steps per run. Capped separately from events, so a long sequential chain is bounded even though it stays narrow. Bundle several items into one step when a chain would otherwise reach five figures.

Concurrency. A wide fan-out is throttled rather than rejected: event creation is rate-limited per run per second, so a flat Promise.all over a few thousand items spends much of its time backing off. Batch or bundle instead — process the list in chunks, or handle several items per step, so fewer and larger units run concurrently. Spawning one child run per item does not by itself narrow the fan-out; it bounds each child's log and isolates failures, which is worth doing for those reasons, but it is not a substitute for chunking.

You cannot raise any of these yourself — WORKFLOW_MAX_EVENTS_OVERRIDE only clamps down, and on the Vercel World the ceilings are service-owned — but Vercel raises the per-run event and step limits on request, so a genuinely large run is a support question as well as a design one.

const BATCH = 100;

async function processItem(item: string) {
  "use step";
  return item.toUpperCase();
}



---

*Content truncated.*