FluxyChat's WorkflowAgent runs long-lived AI agents that survive deploys, restarts, and interruptions. State is persisted to D1/KV, so agents resume exactly where they left off.

WorkflowAgent — Durable Execution

FluxyChat's WorkflowAgent runs long-lived AI agents that survive deploys, restarts, and interruptions. State is persisted to D1/KV, so agents resume exactly where they left off.

Overview

Standard agent invocations are ephemeral — if the Worker restarts mid-run, the agent's state is lost. WorkflowAgent wraps agent execution in a durable layer:

State persistence: Every step, tool call, and intermediate result is saved to D1
Automatic resume: On restart, the agent picks up from the last completed step
Typed context: Shared typed context across steps, tools, and calls
Step boundaries: Each step is atomic — either completes fully or doesn't start

When to Use

Scenario	Use WorkflowAgent?
Quick Q&A (< 5s)	No — standard invoke is fine
Multi-step research (30s+)	Yes — survives timeouts
Code review with tool calls	Yes — durable tool execution
File processing pipeline	Yes — atomic steps
Scheduled recurring agent	Yes — durable scheduling

Basic Usage

import { WorkflowAgent } from "@fluxy-chat/sdk";

const workflow = new WorkflowAgent({
  agentId: "research-agent",
  roomId: "research-room",
  model: "gpt-4o",
  tools: messengerPreset,
});

// Define steps
workflow.step("gather", async (ctx) => {
  const messages = await ctx.searchRoom("pricing discussion");
  ctx.set("findings", messages);
});

workflow.step("analyze", async (ctx) => {
  const findings = ctx.get("findings");
  const analysis = await ctx.llm({
    prompt: `Analyze these messages: ${JSON.stringify(findings)}`,
  });
  ctx.set("analysis", analysis);
});

workflow.step("report", async (ctx) => {
  const analysis = ctx.get("analysis");
  await ctx.postMessage(`## Analysis\n\n${analysis}`);
});

// Run — survives restarts
const run = await workflow.run({ maxSteps: 10 });

Typed runtime context

Define a typed context schema for type-safe data sharing across steps:

import { WorkflowAgent, defineContext } from "@fluxy-chat/sdk";

const ResearchContext = defineContext({
  findings: { type: "array", required: true },
  analysis: { type: "string", required: false },
  confidence: { type: "number", default: 0 },
});

const workflow = new WorkflowAgent({
  context: ResearchContext,
  // ...
});

workflow.step("analyze", async (ctx) => {
  const findings = ctx.get("findings"); // typed as array
  ctx.set("confidence", 0.85);           // type-checked
  // ctx.set("typo", "value");           // TypeScript error!
});

Durable Execution

State is persisted after each step completion:

Step 1: gather    → ✅ persisted
Step 2: analyze   → ✅ persisted
Step 3: report    → 💥 Worker restart
                  → 🔄 Resume from Step 3 (state from Step 2 is loaded)
Step 3: report    → ✅ completed

Multi-step loop control

Configure stop conditions for agent loops:

const workflow = new WorkflowAgent({
  maxSteps: 20,
  stopWhen: (ctx) => ctx.get("confidence") > 0.9,
  // or: stopWhen: hasToolCall("final_answer")
  // or: stopWhen: isStepCount(10)
  // or: stopWhen: isLoopFinished()
});

Tool Execution

Tools called within a workflow are also durable. If a tool call is interrupted:

The tool call is logged in D1 with status: "pending"
On resume, the workflow checks for pending tool calls
If the tool is idempotent, it re-executes
If not, it waits for manual resolution

Monitoring

Track workflow progress via the DevTools UI or API:

const status = await client.getWorkflowStatus(runId);
// {
//   step: "analyze",
//   completedSteps: ["gather"],
//   pendingSteps: ["analyze", "report"],
//   startedAt: "2026-07-01T12:00:00Z",
//   lastCheckpoint: "2026-07-01T12:01:30Z",
//   context: { findings: [...] }
// }

Integration with Cloudflare Workflows

WorkflowAgent integrates with Cloudflare Workflows binding for native durable execution:

export class FluxyScheduledWorkflow extends Workflow {
  async run(event, step) {
    const agent = new WorkflowAgent({ /* ... */ });
    await agent.runWithWorkflow(step);
  }
}

WorkflowAgent — Durable Execution

WorkflowAgent — Durable Execution

Overview

When to Use

Basic Usage

Typed runtime context

Durable Execution

Multi-step loop control

Tool Execution

Monitoring

Integration with Cloudflare Workflows

See Also

On this page