As LLMs became capable of reliably using tools, I realized that this would dramatically simplify the smart contract UX by translating user intent into transactions while also making it possible to embed cheap, flexible cognition directly into the contracts themselves.
AutoTx
My first experiment took place in October 2023, when my team at Polywrap won Safe's AI hackathon with UnblockAI, a chat interface for managing a Safe through natural language. We then pushed further on the agent side with Evo, a general-purpose agent that dynamically chooses the tools it needs to complete a task. It became the #1 performing agent on the AutoGPT Arena leaderboard.
We brought the two together in AutoTx, a general-purpose transaction agent. A user could describe a goal and AutoTx would coordinate specialized agents to research assets, plan transactions, and assemble the actions into a batch for approval through the user's Safe.
We applied the same approach to more specialized problems. Prediction Prophet researched prediction-market questions and placed bets on its forecasts, while Fund Public Goods evaluated projects and proposed how to allocate a donation. For Optimism's Superfest, SuperAgent turned a user's budget, chains, and wallet into a guided set of transactions for completing the festival's onchain missions.
AI-driven DAOs
My chapter for DAOs in the Legal Landscape developed a framework for AI involvement in DAOs across two functions, decision-making and execution, and three levels of autonomy, from AI-assisted to AI-dominant to AI-only. The chapter used this framework to explore what increasingly autonomous organizations could look like, as well as the technical and social limits of delegating their operation to AI.
Edward Elgar Publishing