Research Agent Workflow — Resonate + Google Cloud Functions

This repository contains a single long-running, resumable workflow: a recursive research agent that uses the OpenAI chat completions API and is orchestrated with Resonate. It is designed to run on a serverless platform (e.g., Google Cloud Functions) while the Resonate Server persists and resumes workflow state between spans.

---

Overview

The research workflow implements a recursive research agent:

• It sends a system+user prompt to an LLM (via OpenAI.chat.completions.create).

• If the model returns tool calls asking for sub-research, the workflow:

  • spawns parallel child research runs (ctx.beginRpc) for each subtopic,
  • awaits their results, and
  • feeds results back into the parent message stream as tool outputs.

• If the model returns a plain summary (no tool calls), the workflow returns that summary.

Because the workflow is written as a generator and executed with Resonate, each network call or multi-step operation can be a suspend/resume point — ideal for serverless.

---

Architecture

┌────────────────────────────┐
│  Developer / CLI           │
│  resonate invoke research  │
└─────────────┬──────────────┘
              │
              ▼
┌──────────────────────────────┐
│      Resonate Server         │
│  - persists run state        │
│  - schedules & resumes runs  │
│  - invokes target function   │
└─────────────┬───────────────┘
              │
              ▼
┌──────────────────────────────┐
│ Google Cloud Function (this) │
│  - registers `research`      │
│  - calls OpenAI & orchestrates subruns
└─────────────┬───────────────┘
              │
              ▼
┌──────────────────────────────┐
│ External APIs (OpenAI)       │
└──────────────────────────────┘

---

Environment variables / Secrets

• OPENAI_API_KEY — OpenAI API key (required)

---

Workflow: key implementation notes

• The LLM call is performed via a helper function prompt(...) that returns the model message (and any tool_calls).

• The workflow inspects message.tool_calls and — for each research tool call — uses:

  • yield* ctx.beginRun(research, subtopic, depth - 1) to spawn a child run, then
  • yield* handle to await its result before continuing.

• The generator returns the final text (summary) once the model stops issuing tool calls.

This pattern allows parallel, resumable sub-research tasks that are resilient to process restarts.

---

Triggering & Running

Once the Resonate Server and your Cloud Function are deployed and reachable:

resonate invoke research.<run-id> \
  --func research \
  --arg "causal inference in neural nets" \
  --arg 1 \
  --server https://resonate-server-... \
  --target https://your-cloud-function-url

• research.<run-id> — unique run/promise id of your choice.
• --arg <topic> — topic string.
• --arg <depth> — integer; depth > 0 enables tool/sub-research spawning.