Cloudflare Workflows for Agents

Pair an agent with a durable Cloudflare Workflow for long-running tasks with retries and recovery.

Sby Skills Guide Bot
DevelopmentIntermediate
108/1/2026
Claude CodeCursorWindsurfCopilotCodex
#cloudflare#workflow#agents#durable-execution#typescript

Recommended for


name: ayjnt-workflows description: Pair an agent with a durable Cloudflare Workflow. Use when the user asks to "add a workflow", "run a long-running job", "extend AgentWorkflow", "make order processing durable", "retry with backoff", or "survive worker restarts". Creates agents/<route>/workflow.ts next to agent.ts; the framework wires the workflow binding, the typed RPC stub, and the GeneratedEnv field. Uses the withWorkflow<typeof MyWorkflow>()(Agent) mixin so the agent gets a typed this.workflow(params) shortcut with no magic binding strings. NO migrations — workflows are ephemeral execution containers; state lives in the paired Agent's DO.

Pair an agent with a workflow

workflow.ts next to agent.ts is the entire convention. The default-exported class extends AgentWorkflow (typed RPC back to the agent) or WorkflowEntrypoint (fire-and-forget). The framework discovers it at scan time and emits the binding.

File shape

// agents/<route>/workflow.ts
import { AgentWorkflow } from "agents/workflows";
import type { AgentWorkflowEvent } from "agents/workflows";
import type { WorkflowStep } from "cloudflare:workers";
import type MyAgent from "./agent.ts";

type Params = { /* user-defined params */ };

export default class MyWorkflow extends AgentWorkflow<MyAgent, Params> {
  async run(
    event: Readonly<AgentWorkflowEvent<Params>>,
    step: WorkflowStep,
  ): Promise<unknown> {
    const params = event.payload;

    await step.do("step-one", async () => {
      // durable work
      await this.agent.markStatus(params.id, "running");
    });

    const result = await step.do("step-two", async () => {
      // each step.do persists its return value
      return { something: "useful" };
    });

    await step.sleep("wait-a-bit", "5 seconds");

    return result;
  }
}

Rules

  • Default-export the class. ayjnt scans agents/**/workflow.ts files and reads the default export.
  • extends AgentWorkflow for typed RPC back to the agent. extends WorkflowEntrypoint when the workflow doesn't talk to an agent. ayjnt detects either.
  • Generic params: AgentWorkflow<AgentType, Params> — first arg is the paired Agent class (gives you this.agent with typed RPC); second is the workflow's payload type.
  • No aliased imports for the base class. The detection is source-level regex on extends AgentWorkflow / extends WorkflowEntrypoint.

Triggering from the agent

Use the withWorkflow mixin so you don't have to type the binding name. The framework knows the pairing from the file layout and injects the binding onto the agent's prototype at build time:

// agents/<route>/agent.ts
import { Agent } from "agents";
import { withWorkflow } from "ayjnt/workflow";
import type { GeneratedEnv } from "@ayjnt/env";
import type MyWorkflow from "./workflow.ts";

export default class MyAgent
  extends withWorkflow<typeof MyWorkflow>()(Agent)<GeneratedEnv, State>
{
  override async onRequest(req: Request): Promise<Response> {
    const params = await req.json();

    // No magic binding string — params type-checked against
    // MyWorkflow's Params generic. IntelliSense knows the shape.
    const workflowId = await this.workflow(params);

    return Response.json({ workflowId });
  }

  // The workflow calls this back via this.agent.markStatus(...)
  async markStatus(id: string, status: string): Promise<void> {
    this.setState({ /* … */ });
  }
}

Why curried

withWorkflow<W>()(Base) curries: the outer call binds the workflow type, the inner takes the base class. TypeScript can't partially infer one generic from a type arg and another from a value arg in a single call, so currying is the cleanest way to keep the syntax ergonomic while preserving Agent's <Env, State> parameterization downstream.

For non-co-located workflows

If a workflow lives outside the agent's folder (e.g. a fire-and-forget batch under a separate workflows/ tree), don't use the mixin. Call the SDK's runWorkflow directly with the binding name:

const workflowId = await this.runWorkflow("CLEANUP_JOB", params);

this.runWorkflow(binding, params) is provided by the SDK's Agent base class. The binding is type-checked against bindings ayjnt generated.

Name derivation

Class name drives everything:

  • OrdersProcessing → binding ORDERS_PROCESSING
  • OrdersProcessing → workflow name orders-processing (kebab)
  • OrdersProcessingclass_name: "OrdersProcessing"

Bindings must be unique across agents and workflows; the scanner errors at build time if two classes would collide.

What gets generated

// .ayjnt/dist/wrangler.jsonc
"workflows": [
  {
    "name": "orders-processing",
    "binding": "ORDERS_PROCESSING",
    "class_name": "OrdersProcessing"
  }
]
// .ayjnt/env.d.ts
export type GeneratedEnv = {
  ORDERS_AGENT: DurableObjectNamespace<OrdersAgent>;
  ORDERS_PROCESSING: Workflow;  // ← workflow binding, typed
};
// .ayjnt/dist/entry.ts
import OrdersProcessing from "../../agents/orders/workflow.ts";
export { OrdersProcessing };  // Cloudflare runtime registers the class

Workflows ≠ Durable Objects

  • No entries in migrations[]. Workflows are ephemeral execution containers. The framework will refuse to add a workflow class to the SQLite migrations.
  • State lives in the paired Agent's DO. Workflows don't have this.state. If you need state, RPC back into the agent and have it call setState.

step.do contract

Each step.do("name", fn) block is:

  • Idempotent across retries — if fn throws, the framework retries with exponential backoff. If the worker restarts, execution resumes from the last successful step using the persisted return value.
  • Keyed by name. Renaming a step in a deployed workflow makes the runtime think it's a new step and re-execute it. Treat step names like database column names — stable across deploys.

step.sleep("name", duration) also survives restarts.

Pair with a UI

The agent ⇄ workflow loop is naturally visible — each step.do flips the order's status, which broadcasts over WebSocket. Drop an app.tsx next to the agent and use the generated useAgent hook:

// agents/<route>/app.tsx
import { useAgent } from "@ayjnt/orders";

export default function OrdersUI() {
  const agent = useAgent();
  const orders = (agent.state as { orders?: Order[] })?.orders ?? [];
  // call agent.call("placeOrder", [{ sku, qty, customerId }]) on submit
  // render orders with their live status
}

Make the agent's trigger method @callable({ description: "..." }) so the UI can drive it via agent.call(method, args) without going through fetch.

Looking up a running workflow

// Inside the agent
const instance = await this.env.MY_WORKFLOW.get(workflowId);
const status = await instance.status();
// → { status: "running" | "complete" | "errored" | ..., output?: ... }

When NOT to use a workflow

  • The task is a single fetch + return → just do it in the agent.
  • The task is a stateless background side-effect → use{" "} this.ctx.waitUntil(fetch(...)).
  • The task needs millisecond-precision orchestration → workflows have per-step overhead; use the agent's scheduler instead.

Reach for workflows when you want durable, multi-step, retry-aware execution.

Reference

Related skills