DEV Community

Zoe Lin
Zoe Lin

Posted on

Email Drafter — Multi-Agent Email Writing with Google ADK and Cloud Run

#agents #buildmultiagents #gemini #adk

Education Track: Build Multi-Agent Systems with ADK

What I Built

I built Email Drafter, a small web app that turns structured user input into a polished email draft using a multi-agent workflow.

The app takes:

  • recipient type
  • purpose
  • tone
  • language
  • key points

Instead of using one giant prompt, I split the workflow into multiple specialized roles.

The system works like this:

  • one agent plans the email
  • one agent reviews the plan
  • one agent writes the final draft
  • one orchestrator coordinates the workflow

I also deployed the system as separate services on Cloud Run, so the app feels much closer to a real distributed AI workflow instead of just a local prototype.

I picked email drafting because it is a practical use case where planning, reviewing, and writing feel like naturally separate tasks. That made it a good fit for experimenting with multi-agent orchestration.

Cloud Run Embed

Your Agents

This project uses three specialized agents plus one orchestrator.

  • Researcher Agent

    Creates a structured email plan from the user input.

  • Judge Agent

    Reviews the plan and checks whether it is complete enough before moving on.

  • Content Builder Agent

    Writes the final email draft from the approved plan.

  • Orchestrator Agent

    Coordinates the workflow between the other agents and returns the final result to the frontend.

The overall flow is:

Frontend -> Orchestrator -> Researcher -> Judge -> Content Builder -> Final Email Draft

Key Learnings

This project helped me understand why multi-agent systems can be useful even for a relatively small app.

A few things stood out while building it:

  • Splitting the workflow made debugging easier. It was much easier to tell whether the issue came from planning, review, or writing than trying to fix one large prompt.
  • Deployment was harder than the basic prompt logic. Getting multiple services running locally, wiring them together, and then deploying them to Cloud Run took more effort than writing the first version of the agents.
  • Prompt wording had a big impact on output quality. Small instruction changes made a noticeable difference between getting a planning-style response and getting something that looked like a real email.
  • Multi-agent design felt more modular and production-oriented. Even for a simple email drafting tool, separating responsibilities made the system easier to reason about and improve.

Demo

Top comments (0)