Agents Rap Battle: Realtime AI Performance Experiment

Performers, not autonomous tool agents.

The word agents can suggest a tool-using system, but that is not the claim here. These are persona-driven performers: two generated voices with defined roles that share a conversation context and take turns inside a designed performance loop.

• The experience is about coordination and character.
• There are no scale or autonomy claims attached to the demo.
• The product surface makes the performance legible as a shared room.

One room, shared context.

A Cloudflare Worker handles the edge entry point while a Durable Object gives each session a place to coordinate state. WebSockets keep the browser connected to the room so the next turn can be streamed instead of waiting for a full page refresh.

• Durable Object state gives a session a clear owner.
• WebSocket messages keep turn-taking visible.
• The system makes the handoff between performers part of the interaction.

Gemini meets Web Audio.

Gemini generates verses and performer-specific speech while Web Audio handles the browser's beat-aware playback. Bilingual verses, rhyme and punchline cues, and a live score display keep the audience oriented. Audio is cached in R2 for replay, and a recap scorecard supports sharing or starting a rematch.

• Streaming verses create a sense of momentum.
• Audio feedback gives the browser experience a physical cue.
• Persisted room state supports reconnecting to a battle instead of losing the session.

Can coordination feel like a show?

The experiment tests a narrow question: whether shared context, turn-taking, and a little stage direction can make AI output feel more like an event. It is a focused prototype, not evidence of production scale or a general agent platform.