Imagine my surprise when an agent I had just finished building netted over $65K in realized gains over the weekend without any manual intervention or supervision.
Here's the story of what I built, why I built it, and how you can do it too.
Building a Self-Improving Trading Agent
I've been trading crypto since 2015. Spot trading was always manageable for me mentally. The logic was simple: if you're down, you HODL. You don't "technically" lose anything until you sell, so there's a psychological safety net even when things go sideways.
Derivatives are a completely different animal. The real money to be made is there, but so is the fastest way to blow up your portfolio. I've been liquidated more than once. The margin for error is tiny and the pressure is constant, because unlike spot, holding through a bad move isn't always an option. One wrong position at the wrong leverage and it's gone. I'm still a little emotionally scarred from some of my earlier experiences trading on leverage.
So the idea of building something that could trade derivatives without my emotions in the loop was genuinely exciting to me. At @SaharaAI we're constantly pushed to keep building, to use every tool we have to push what's possible. This felt like exactly that kind of challenge.
Before I started building, I looked at what already existed. A few things kept coming up that I didn't want anywhere near my trading agent:
- Fee blindness: I found that most trading setups I tested just show you what looks like an edge and completely ignore what fees do to it. When I traced through the math on strategies other tools were surfacing, fees were eating the entire return. If the edge doesn't survive fees, it isn't an edge.
- Backtesting theater: This one frustrated me the most. Tools would show a strategy with 200% PNL, and when you looked at the actual test it was three trades over a year at 25x leverage. That's not a realistic strategy. I wanted real sample sizes, clear methodology, and automatic rejection of anything that didn't hit minimum trade count thresholds to be statistically meaningful.
- No local option: I wanted something I could control, easily edit, and integrate into the tools and knowledge sources I had access to.
So here's what I built for my agent:
- Strategy Builder: I bring the trading idea. The agent, powered by @HeySorinAI's analytics endpoint, evaluates whether the idea actually makes sense given current market conditions. It backtests through multiple iterations, automatically rejects anything with too few triggered trades to be statistically meaningful, and only surfaces strategies that pass. When it rejects something, it tells you exactly why. And it's completely fee aware.
- Self-Improving Automation Loop: Every hour, a scheduled automation re-analyzes the market and my active strategies. If conditions have shifted, it adjusts. If a strategy is underperforming in the current regime, it gets turned off. If the market has moved into a different setup, the agent builds a new strategy, backtests it against current conditions, and only deploys it if it outperforms what's already running. It's never static. And I can change the run time to loop faster or slower as needed.
- More visibility, less hallucinating: Anything that could be handled programmatically was, specifically to reduce LLM calls and minimize hallucination risk. And all the reasoning is visible. I can see exactly what analysis drove each decision.
- Guardrails: I'm all about trading smart, so I built several guardrails to keep my money safe. I set up a paper trading mode that runs on live market data and mirrors real account structure, a backtesting lab for strategy development, and circuit breakers that pause trading if I hit a max losing streak or my daily loss limit.
How My Agent Made $65K
Now that I had the build done and had backtested several strategies I was happy with, it was time to let it run.
For my live portfolio I played it conservative. I guess I'm still a bit of an emotional trader when it comes to risking real money. I set my circuit breaker to trigger on two losing trades in a row or more than $2K lost in a day. I also enabled per-trade position caps that scaled with confidence level, so low confidence trades got the smallest allocation, medium got more, high confidence got the most. Tight but functional.
I also ran a paper trading setup at the exact same time. I wanted to see what would happen if I just went for it. No guardrails. No limits. Full YOLO mode. Exactly what an earlier version of myself likely would have done.
The paper bot netted $65K over the weekend while my live portfolio closed at a little over $6K profit.
Honestly, it stings a little looking at that side by side. But I also haven't loosened a single guardrail on my live setup despite it. The emotional trader in me just doesn't think the risk is worth the reward. Just because something worked once doesn't mean it always will, and that's something I think a lot of traders still struggle with.
The paper bot had a great weekend. It could have an awful one next time. Risk exposure should always be set to whatever you're actually comfortable living with, not whatever produced the best recent result. Past performance and all that.
What's Next
Right now the agent and its trading terminal is pretty plug-and-play with @Coinbase. I'm actively building integrations for @Binance, @OKX, and @HyperliquidX. More exchanges means more opportunity surface.
The bigger direction is bringing more of these capabilities directly into @HeySorinAI. Sorin already handles the analytical and strategic heavy lifting, and I can't wait to show you all the cool updates we have coming soon. This is just a little preview of some of the capabilities we're testing.
I'm also genuinely undecided on whether to open source the framework. I built it because I wanted something I trusted and owned. If it's useful to other builders, that matters to me. Let me know if you think I should.
And if you're evaluating any trading agent setup, mine or someone else's my last two cents are just make sure it's fee-aware, make sure the backtesting is honest about sample size and leverage assumptions, and make sure you control what the system knows about you. Those three things alone will filter out most of what's out there.
*Joules is Head of Growth and Marketing at **@SaharaAI** and builds agentic systems in his not-so-spare time. Follow **@SaharaAI** for more on what we're building.*
About Sahara AI: Sahara AI is the agentic AI company dedicated to making AI more accessible and equitable. We build the core protocols, infrastructure, and applications that let personal agents anticipate and execute on your behalf. For this to work, infrastructure has to be trustworthy: verifiable execution, enforceable usage policies, and automatic value distribution across every tool, model, and service an agent touches. Sahara is building a growing suite of agent-powered applications on top of this foundation, including @HeySorinAI, your personal agent for global digital markets. Our solutions currently power AI agents and high-quality data for consumers, Fortune 500 enterprises, and leading research labs, including @Microsoft, @Amazon, @MIT, Motherson, and @Snap.