Harshil, a B.Tech AI/ML student in Delhi, along with teammates Daksh and Kartik, built Delegates Courtroom for the Flip the Switch hackathon. The project addresses a common friction point in creative workflows: when AI agents can suggest compromises, but humans must retain the final veto. The team’s core rule is explicit—agents negotiate, but humans sign.

The Switch Architecture

The system relies on Switch Rooms, which provide a shared context log for both people and agents. Unlike traditional setups where one assistant’s output is copied into another’s prompt, Switch agents work from the same message history. The team used Mattermost to bridge the room to a chat interface, allowing users to send decisions as simple chat messages. The infrastructure is pinned to Switch core version 0.31.0 and Mattermost 11.9.0, running on a local Docker stack with Node.js 22.13 or newer.

Negotiating Without Consent

In the demo scenario, a creative lead wants a portrait for a promo, but the creator refuses to have their face used. Two AI advocates, powered by Codex CLI, negotiate a storyboard based on conflicting briefs. A deterministic Node.js clerk, not a model, enforces the rules. During a recorded rehearsal, the lead’s advocate proposed a board that violated the creator’s brief. The creator’s advocate caught the conflict and requested an amendment, pausing the process for an owner decision.

Infrastructure and Identity Management

Setting up the environment required careful identity management. The team ran Switch Console in WSL using a Linux .deb package, but the background-worker helpers were written for Windows PowerShell. A critical lesson from the build was tying human accounts to owners. Because the pinned HTTP directory didn’t expose native-account mapping, the team used a scoped, read-only SQL query to verify sender identity, ensuring that only trusted users could approve briefs.

The Human-in-the-Loop Reality

It is important to note that the initial rehearsal used synthetic fixtures for owner decisions. While the advocate turns went through real Switch and Codex instances, the approvals were simulated. The output is an HTML storyboard and text script, not a rendered video. The team emphasizes that passing tests—113 in a fresh clone check—does not equate to real human consent. The goal now is to record a workflow where both real owners review and approve the same bundle.

Key Takeaways

  • Give participants unique identities, not just names, to ensure agents respond to the correct owner.
  • Make waiting a valid state; if an amendment needs an owner, the models must stop rather than hallucinate approval.
  • Separate infrastructure issues from application logic, such as the Docker socket error encountered during setup.
  • Recognize that passing technical tests is different from securing real-world human consent.

The Bottom Line

The most valuable feature in AI agent orchestration isn't the negotiation—it's the enforced pause. By making consent a hard gate rather than a suggestion, Delegates Courtroom shows how to keep humans in control without slowing down the creative process.