I Made My Claude AI Agents Invent Their Own Language. It Cut Token Costs by 63%.

The author created a TikTok-style video platform using AI agents, initially impressed by their efficiency. However, the author realized the agents' English communication was inefficient, wasting tokens on conversational filler. To address this, the author developed "AgentSpeak," a concise, token-saving protocol for agent communication. AgentSpeak uses status codes, action symbols, and file shortcodes to convey information efficiently, drastically reducing token usage. Real-world benchmarks showed significant token savings, particularly for code-related updates. The hub-and-spoke architecture, read-only QA, and file ownership strategies further minimized conflicts. The author emphasizes the importance of verifying results beyond compiler checks, highlighting the need for lead verification in the actual product. The author provides easy instructions for implementing the protocol and making it open source. Finally, the project includes the AgentSpeak protocol, team setup patterns, and guides for optimization.