The “AI Hands” created by PowerPool won the POND track on the dAGI hackathon
As we build the PowerPool DePIN layer for automation, we’ve noticed that more and more AI Agents dedicated to operating in Web3 are being released. These AI agents use on-chain and off-chain information to make decisions that should be implemented as a state update in some contracts. Examples of such decisions include trading and using various Defi strategies.
However, by default, AI agents cannot execute their decisions as transactions. This is simply not their core business. Some protocols that build AI agents have added an option to run or use bots provided by the protocol to implement decisions.
To highlight this issue, the PowerPool team members created the “AI Hands” project and presented an effective solution to the AI transaction execution challenge. The solution earned a bounty at the POND track during the dAGI hackathon, showcasing traction and receiving positive feedback from AI developers and judges.
This article will provide details of this solution and further development plans.
The AI Hands by PowerPool
AI Hands executes transactions on behalf of AI Agent users. We built it on top of POND, a protocol focused on providing trading decisions to its users based on cluster analysis of on-chain data.
The AI Hands is an off-chain resolver built on top of the PowerPool Keeper network. It listens to the POND API and executes transactions on behalf of the trading agent user when certain conditions are met.
The algorithm has a single decision resolver function and utilizes PowerPool’s Keeper network. It begins by calling the POND API to obtain model metrics. Trend accuracy is chosen as the target metric because the hourly forecast interval allows for linear approximation.
A typical off-chain resolver in PowerPool works as follows:
The algorithm includes safeguards to ensure the reliability of its decisions. If the accuracy of the current model falls below a certain threshold, the results are discarded and the execution instance is terminated. Conversely, if the accuracy exceeds the threshold, the POND API is called again to obtain predictions for the unstable token in the pair and the current price.
To further refine the decision process, the algorithm evaluates the predicted change of the unstable token in relative units. If this change falls below a user-defined threshold, the results are discarded again, as minimal changes may not justify the risk or cost associated with a swap.
The overall scheme of AI Hands operations is presented below:
The code is available here.
Conclusions and Future Plans
The execution of AI-generated transactions is an emerging market with huge growth potential.
Use cases include AI-resolved intents, defi strategies, trading strategies, and many more. We plan to integrate POND and provide its users with an easy onboarding option to outsource transaction execution to PowerPool. Other protocols, especially intent-based, will follow, enriching the PowerPool ecosystem with a set of use cases and templates that provide users with a one-click execution of decisions generated by AI agents.
