Inspiration
I’ve always been interested in AI that can actually learn from mistakes instead of just following static prompts. Most agents don’t improve when they mess up—I wanted to change that. Inspired by how humans learn to trade, I set out to build an agent that reflects, adapts, and gets better over time.
What it does
MAVERICK is an AI trading agent that uses real-time data to make BUY or SELL decisions on cryptocurrencies. After each trade, it evaluates whether the decision was profitable. If not, it reflects on what went wrong and rewrites its own system prompt to improve its future decisions.
How we built it
We used Claude’s Agent SDK and MCP servers to give MAVERICK access to live data via Brave Search (news/sentiment) and CoinGecko (market data). The agent makes a decision, waits 30 seconds to evaluate the price movement, and then either reinforces or adjusts its strategy based on the outcome. A custom web UI helps visualize this loop in real-time.
Challenges we ran into
Getting the feedback loop to actually work was tough. Early versions made random or repetitive decisions. It took a lot of iteration to get Claude to meaningfully analyze failures and produce better prompts. Finding the right timing window (30 seconds) for fast but useful evaluation also required experimentation.
Accomplishments that we're proud of
We built an AI agent that not only trades but learns and adapts after each mistake. Seeing the prompt evolve and the strategy improve over multiple runs was a big win. The live visualization makes the process transparent and insightful.
What we learned
Self-improvement in AI isn’t automatic—it has to be designed. The quality of reflection, access to the right context, and how that feeds into future behavior all matter. Iterative prompt engineering is powerful when done right.
What's next for MAVERICK
We want to expand MAVERICK’s timeframes, improve its reflection methods, and test it on more volatile conditions. We’re also exploring ways to apply this self-correcting loop to other decision-making domains beyond trading.
Log in or sign up for Devpost to join the conversation.