Deep Research Agent on Cloudflare Workers

A Research Agent powered by Resonate and OpenAI, running on Cloudflare Workers. The Research Agent is a distributed, recursive agent that breaks a research topic into subtopics, researches each subtopic recursively, and synthesizes the results.

How It Works

This example demonstrates how complex, distributed agentic applications can be implemented with simple code in Resonate's Distributed Async Await: The research agent is a recursive generator function that breaks down topics into subtopics and invokes itself for each subtopic:

function* research(ctx, topic, depth) {
  const messages = [
    { role: "system", content: "Break topics into subtopics..." },
    { role: "user", content: `Research ${topic}` }
  ];

  while (true) {
    // Ask the LLM about the topic
    const response = yield* ctx.run(prompt, messages, ...);
    messages.push(response);

    // If LLM wants to research subtopics...
    if (response.tool_calls) {
      const handles = [];

      // Spawn parallel research for each subtopic
      for (const tool_call of response.tool_calls) {
        const subtopic = ...;
        const handle = yield* ctx.beginRpc(research, subtopic, depth - 1);
        handles.push([tool_call, handle]);
      }

      // Wait for all subtopic results
      for (const [tool_call, handle] of handles) {
        const result = yield* handle;
        messages.push({ role: "tool", ..., content: result });
      }
    } else {
      // LLM provided final summary
      return response.content;
    }
  }
}

The following video visualizes how this recursive pattern creates a dynamic call graph, spawning parallel research branches that fan out as topics are decomposed, then fan back in as results are synthesized:

call-graph.mov

Key concepts:

• Concurrent Execution: Multiple subtopics are researched concurrently via ctx.beginRpc
• Coordination: Handles are collected first, then awaited together (fork/join, fan-out/fan-in)
• Depth control: Recursion stops when depth reaches 0

---

Running the Example

You can run the Deep Research Agent locally on your machine with Cloudflare's Wrangler or you can deploy the agent to Cloudflare Platform.

1. Running Locally

1.1. Prerequisites

Install the Resonate Server & CLI with Homebrew or download the latest release from Github.

brew install resonatehq/tap/resonate

To run this project you also need an OpenAI API Key and export the key as an environment variable

export OPENAI_API_KEY="sk-..."

1.2. Start Resonate Server

Start the Resonate Server. By default, the Resonate Server will listen at http://localhost:8001.

resonate dev

1.3. Setup the Deep Research Agent

Clone the repository

git clone https://github.com/resonatehq-examples/example-openai-deep-research-agent-cloudflare-ts
cd example-openai-deep-research-agent-cloudflare-ts

Install dependencies

npm install

1.4. Start the Deep Research Agent

Start the Cloudflare Function. By default, the Cloudflare Function will listen at http://localhost:8080.

npm run dev

1.5. Invoke the Deep Research Agent

Start a research task

resonate invoke <promise-id> --func research --arg <topic> --arg <depth> --target <function-url>

Example

resonate invoke research.1 --func research --arg "What are distributed systems" --arg 1 --target http://localhost:8787

1.6. Inspect the execution

Use the resonate tree command to visualize the research execution.

resonate tree research.1

2. Deploying to Cloudflare

This section guides you through deploying the Deep Research Agent to Cloudflare Platform using Cloud Run for the Resonate server and Cloud Functions for the research function.

2.1 Prerequisites

Resonate

Install the Resonate CLI with Homebrew or download the latest release from Github.

brew install resonatehq/tap/resonate

OpenAI

To run this project you need an OpenAI API Key.

Cloudflare Platform

Ensure you have a Cloudflare Platform account.

Warning

Cloudflare Platform offers extensive configuration options. The instructions in this guide provide a baseline setup that you will need to adapt for your specific requirements, organizational policies, or security constraints.

2.1 Create a Hello World Worker

2.2 Deploy the Resonate Server to Workers

Expose the Resonate server running locally to the cloud

Step 1: cloudflare tunnel url (you can use ngrok or similar)

cloudflared tunnel --url http://localhost:8001

Step 2: Run the server locally

Configure the Resonate Server with its URL.

resonate dev --system-url <tunnel-url>

Example

resonate dev --system-url https://purple-lion-2q3j4j.trycloudflare.com

2.3 Deploy the Countdown to Workers

Paste the wrangler.toml at the root of this directory (replace the name with the name of your newly created worker)

name = "XXXXX-XXXXX-XXXX"
main = "src/index.ts"
compatibility_date = "2024-01-01"

Deploy it

npm run deploy

2.4 Invoke the Deep Research Agent (Do not forget to set the API key as env variable at cloudflare)

Start a research task

resonate invoke <promise-id> --func research --arg <topic> --arg <depth> --target <function-url>

Example

resonate invoke research.1 --func research --arg "What are distributed systems" --arg 1 --target https://xxxxx-xxxxx-xxxx.abc.workers.dev

2.5. Inspect the execution

Use the resonate tree command to visualize the countdown execution.

resonate tree research.1

research.1
├── research.1.0 🟢 (run)
├── research.1.1 🟡 (rpc research)
│   └── research.1.1.0 🟡 (run)
├── research.1.2 🟡 (rpc research)
│   └── research.1.2.0 🟡 (run)
└── research.1.3 🟡 (rpc research)
    └── research.1.3.0 🟡 (run)

Troubleshooting

If you are still having trouble please open an issue.