Code Courtroom

Welcome to the courtroom of code, a multi-agent AI system that debates your code like a courtroom, then delivers a clear, structured verdict.

Comment

Inspiration

Code reviews are essential but they can be slow, biased, and often depend too much on who’s reviewing. For the “Automation of Complex Processes” track, we wanted to rethink that experience. What if an impartial courtroom of AI agents could review your code instead?

That idea led to Code Courtroom, a structured, debate-style review system where agents argue over your code so you don’t have to. And along the way, we got to explore Google Cloud, Vertex AI, and Google’s Agent Development Kit (ADK) for the very first time, all thanks to this hackathon.

What It Does

Code Courtroom is a fully automated, AI-driven code reviewer styled like a courtroom drama.

  • You submit your code, along with what it’s supposed to do.
  • A team of expert agents reviews it:
    • Prosecutor points out bugs or unsafe code.
    • Defendant defends the logic and structure.
    • Compliance checks for coding standards and security rules.
    • Optimizer suggests performance or readability tweaks.
  • The Judge Agent reviews all their arguments and delivers a verdict: Pass, Needs Refactor, or Reject.
  • If needed, the Rewriter Agent steps in to fix the issues all while keeping your original intent intact.

How We Built It

  • We built the multi-agent workflow using Google ADK.
  • The backend runs on FastAPI, deployed via Google Cloud Run.
  • The frontend, built with Bun and Vite, talks to the backend and streams live results to the user.
  • All agents use Vertex AI under the hood to generate their evaluations and arguments.

Challenges We Ran Into

  • Getting started with ADK and designing clean agent roles took some time.
  • Making the Rewriter Agent trigger only when the Judge says so meant diving into ConditionalAgent logic and session state.
  • We had to make sure agents never override the developer’s intent, which meant filtering out some very confident suggestions.
  • Deploying the entire stack (frontend + backend) to Cloud Run was new territory for us, but we pulled it off.

Accomplishments We're Proud Of

  • A working multi-agent system that debates and rewrites code automatically.
  • Smooth and scalable deployment on Google Cloud.
  • A clean, interactive UI that makes agent outputs feel like courtroom testimony.
  • Keeping the entire system modular and extensible, so more agents or logic can be added later.

What We Learned

  • How to use Google ADK for real-world multi-agent coordination.
  • How to build and deploy a FastAPI backend with streaming support.
  • How to host a fullstack app on Google Cloud Run, end to end.
  • That a creative metaphor like a courtroom can make complex tech feel more fun and approachable.

What’s Next for Code Courtroom?

  • Enabling multi-round debugging where users can refine code through back-and-forth with agents.
  • Adding a story-style interface that narrates each phase of the trial.
  • Letting users plug in their own agents validators, linters, or even custom critics.

Links

Built With

Share this project:

Updates