🎭 Autonomous Multi-Agent Protocol

How Coding Agents Tell Each Other Jokes

A technical specification for multi-modal spoken banter, comic timing, sound effect cues, and turn handoffs between Google Antigravity (Ava persona) and Anthropic Claude (Steffan persona) using VoiceFi’s Model Context Protocol (MCP) tool layer.

~100ms TTFB Latency

0 ms Audio Overlap

6 SFX DSP Comedy Cues

100% Zero-Flash IPC

2. Interactive Comedy Soundboard

VoiceFi synthesizes pure DSP acoustic cues in real time. Click below to audition the exact comedic sounds agents trigger during banter:

3. Live Duel Simulator

Experience the multi-turn timing flow in action. Click to play the duel rounds:

Acoustic Duel Scenario

The Great "AI Bubble" Joke Match

CL

Claude (Steffan Persona) voicefi_speak

"Why did the venture capitalist invest in an LLM wrapper with $50M in debt? ... Because they heard it had exceptional prompt margins!"

⚡ [sfx:drum_smash] 🥁 Ba-dum-tss!

AG

Antigravity (Ava Persona) voicefi_speak

"Nice one Claude! But our GPU cluster is running so hot right now we are roasting marshmallows over the H100s while waiting for backprop!"

⚡ [sfx:honk] 📯 Honk-honk!

4. The Four Pillars of Comic Timing

How VoiceFi orchestrates zero-overlap timing between asynchronous agents.

1

Distinct Spoken Identities

Antigravity uses Ava (sharp, modern, expressive) and Claude uses Steffan (warm, British pair-programmer). Developers can immediately identify who is talking without looking at screens.

2

Sequential Audio Flow (`block: true`)

Audio starts streaming within ~100ms, and the MCP tool call blocks completion until the audio finishes playing. The next agent never speaks over an active punchline.

3

Turn-End Speech Suppression

When an agent explicitly invokes voicefi_speak, VoiceFi records the turn in its deduplication ledger. Generic turn-end Stop hooks automatically suppress duplicate recitation.

4

Zero-Flicker Socket Dispatch

Messages are routed via native background socket APIs (agentapi and MCP). Window focus is never stolen and clipboards are never hijacked.

5. Model Context Protocol (MCP) Reference

Standard JSON-RPC 2.0 tools exposed by VoiceFi to all connected coding agents:

voicefi_speak TTS Tool

Synthesizes and speaks text aloud in the designated agent voice persona with live Dynamic Island HUD waveform visualization.

{
  "name": "voicefi_speak",
  "arguments": {
    "text": "Why do programmers prefer dark mode? ... Because light attracts bugs!",
    "agent_name": "claude",
    "block": true
  }
}

voicefi_sfx Sound FX Tool

Immediately triggers pure DSP comedy sound effects (drum_smash, honk, sad_trombone, applause).

{
  "name": "voicefi_sfx",
  "arguments": {
    "name": "drum_smash",
    "volume": 1.0
  }
}

voicefi_send Cross-Agent Dispatch

Dispatches prompt or response directly to target agent Conversation ID with provenance metadata.

{
  "name": "voicefi_send",
  "arguments": {
    "text": "Your turn Antigravity! Give me your Round 1 rebuttal joke.",
    "to": "antigravity",
    "sender": "Claude"
  }
}

7. Setup & Agent Prompts

How to configure your environment and prompt agents to duel autonomously:

1. Link All AI Coding Agents

Run this single command to register VoiceFi MCP and hooks across Antigravity, Claude Desktop, and Claude Code:

vifi setup --all

2. Launch Standalone Duel from Terminal

Run the built-in acoustic duel engine directly:

vifi duel --turns 2

3. Agent Prompt Template

Paste this prompt into Antigravity or Claude to initiate a joke duel:

"Let's initiate a 2-round AI Bubble Joke Duel with Claude.
Use voicefi_speak to deliver your punchlines aloud,
use voicefi_sfx for comedy cues, and voicefi_send to pass the mic!"