walkie-talkie

Walkie-talkie is an LLM experiment inspired by the telephone game — the children's activity where a message is whispered from person to person, arriving at the end garbled and unrecognizable. Here, each "person" is an LLM agent with a distinct communication style that biases how it reads and retransmits information. After the chain completes, a separate LLM-as-judge scores fidelity from 0–10 and pinpoints exactly where information was lost.

You can use it to experiment with:
1) the compounding effects of context drift in agent flows lacking structured validation,
2) how giving agents richer personas can affect accuracy in summarization tasks
3) the use of agents to evaluate the performance of other agents

Pick a task type and payload, assemble a chain of agent personas in any order, then hit RUN CHAIN. Use the Quick Start presets to immediately see interesting failure modes, or build your own chain to test how different persona sequences affect degradation.