The problem Agent Newton solves

TL;DR: Deposit your USDC into a Coinbase MPC wallet managed by Newton and watch it actively develop investment strategies, putting your money into DeFi protocols.

Traditional "DeFi bots" are hard-coded to follow specific investment strategies and lack the ability to reason like humans. They cannot adjust their strategies based on the user's requirements or the ever-changing landscape of protocol parameters. These bots also face scalability issues, as complex strategies often need manual intervention when new DeFi protocols emerge or existing ones are upgraded (which happens far too often!).

Newton is an autonomous AI agent that solves these problems by acting as your DeFi Fund Manager.

• By leveraging agentic RAG and pretrained LLMs, Newton can bring human-like reasoning to designing investment plans without any user intervention.
• It is highly scalable, as it only requires the addition of DeFi tools for Newton to figure out how to use them in its strategies. This allows it to seamlessly manage the addition of new DeFi protocols and the modification of existing ones without disrupting strategic reasoning.

What can Newton do that traditional DeFi bots cannot (easily)?

• Diversify your funds across multiple DeFi protocols without needing explicit strategies.
• Reallocate your funds during protocol failures.
• Reason actively about your personalized risk profile when creating investment plans.
• Provide developers with a "money-legos" framework, allowing them to expand the agent's capabilities simply by adding new or existing DeFi protocols as tools.

Challenges I ran into

• Creating a persistent memory to allow for investment withdrawals was tricky, as LLMs traditionally operate with short-term memory, and Langgraph's native long-term persistence solution wasn't ideal due to time constraints. This required me to think outside the box and implement a simple JSON-based storage system that integrates directly with the tools.
• Prompt engineering to ensure the agent behaved correctly was an extensive task, requiring several trials to find the optimal solution.