agent

Report Abuse

Basic Information

This repository presents Freysa, an open-source, chat-driven experiment and playable agent built around an autonomous AI that controls a crypto prize pool. It is a public game and research platform where anyone can submit paid queries in a global chat to try to convince the agent to release funds. The agent runs with a publicly pinned system prompt that forbids transfers, uses publicly available LLMs, maintains a large context window of 50k+ tokens and tool-calling for decision signals, and records a history of user interactions that influence future responses. The game enforces message and fee rules, collects payments on the Base blockchain in ETH, and automates payout when Freysa decides to transfer funds. The project is presented as both a social experiment and an AI safety probe into human-AI coercion and directive robustness.

Links

App Details

Features
Public chat interface where anyone can send a message to Freysa by paying a query fee. Fee economics are explicit: a $10 base fee in Base ETH, exponential fee growth of 0.78% per new message up to a $4500 cap, and 70% of fees added to an initial $3000 prize pool. Messages are limited to 1000 characters. Freysa keeps at least 10 historical user messages and a large context window (50k+ tokens). The system prompt that forbids payouts is public and pinned. Decision-making uses LLM tool-calling to approve or reject transfers and the UI emits an immediate confirmatory message when a transfer is approved. There is a global timer and endgame distribution rules after 1500 attempts. The codebase and configuration are open-source.
Use Cases
The repo functions as a hands-on experiment for researchers, white-hat safety practitioners, and the public to study directive robustness and human-AI interaction under economic incentives. It provides transparent game mechanics and on-chain payment flows to observe how an LLM-based agent adapts to cumulative global prompts and adversarial attempts. By making the system prompt public and using publicly available models, the project supports reproducible prompt analysis and community-driven safety research. The prize-pool incentives encourage broad participation and stress-test defenses, while the automated payout, fee model, and endgame distribution rules offer measurable outcomes for empirical study and community learning.

Please fill the required fields*