I Launched Automated AI Stock Trading Agents 5 Days Ago. Here’s What I Learned.

Lessons From Creating a Free No-Code AI Agent for Stock Trading

8 min readOct 11, 2025

Five days ago, I launched Aurora 2.0.

In other words, I turned a boring chat bot into a powerful AI Agent.

AI Stock Trading Agent

Use a fully autonomous AI agent to perform financial research and create algorithmic trading strategies.

nexustrade.io

Unlike general-purpose Large Language Models, these agents have highly specialized tools to allow you to build personalized trading strategies. I launched this feature exactly 5 days ago and over 270 agents have been created so far.

What happened next completely changed how I think about AI agents.

TL;DR:

Launching A Truly Revolutionary Stock Trading Agent

For context, I’ve been working on NexusTrade since I was a student at Carnegie Mellon and getting my Masters degree. For the past 5 years, I’ve been adding features, iterating on the design, and building out a no-code platform for creating trading strategies.

The standout feature was an AI chatbot. It could take requests like “build me a trading strategy to rebalance the Magnificent 7 every two weeks”, and transform that into a strategy where you can update, backtest, optimize, and deploy.

But I didn’t stop there.

The New NexusTrade AI Agent can autonomously create, backtest, optimize, and deploy trading strategies

Taking lessons from Claude Code and Cursor, I transformed my boring chat into fully autonomous AI agent.

And the lessons in these five short days have been WILD.

Want to use AI to build your trading strategy? NexusTrade’s AI Stock Trading Agent is free for a limited time!

1) AI Agents Are WAY More Expensive Than You Think

My Dashboard for Requesty — I can spend $60+ per day on agents

I’ve gained a newfound respect for the Cursor and Claude Code teams.

And their accounting department.

AI Agents are expensive. Very expensive. Even when using an inexpensive but capable model like Gemini 2.5 Flash, which costs $0.30/M input tokens and $2.50/M output tokens, the cost of calling external tools, retry logic, and orchestration is exorbitant, to the point where I’m paying $60+ per day on these agentic functionalities.

However, let me make my confident prediction right now – this will NOT be an issue 1 year from now.

The cost of models have been decreasing rapidly while they're capabilities have gotten better and better. this time next year, we’ll have a model that's more capable than Claude 4 Opus, but costs less than $0.20/M input and output tokens.

I’m calling it right now.

But it wasn’t the insane costs that really made my jaw drop this past week.

No, it was seeing (and understanding) how insanely important prompt engineering ACTUALLY is.

💡 Quick Tip: Want to see exactly how much agent runs cost? View Live Cost Dashboard — Watch real-time token usage by clicking on the purple graph

See agent costs, tool calls, and even gantt charts all with the click of a button!

2) Prompt Engineering is 3x More Important Than You Think

Most failures don’t come from the model — they come from vague prompts.

If you want your agent to actually reason about problems, call tools, and generally unlock REAL insights, you’re probably going to have to spend months refining your prompts.

Prompt engineering is far more important than the tech crowd gives a credit for. A good prompt is the difference between a model being slow and inaccurate vs fast and reliable. Few-shot prompting, clear instructions with no ambiguity, and even retrieval-augmented generation can all help with building an AI agent that can solve very complex tasks.

Such as “how to build a trading strategy”.

For example, my system has over 14 public-facing prompts and 6 internal prompts to make it run autonomously. Each prompt is extremely detailed, often containing:

A detailed description for when to use the tool
Instructions on what to do and what NOT to do
A schema that the AI should adhere to when responding
Few-shot prompting examples that shows the AI how to respond

The left-hand side shows the instructions, the right hand side tells the Agent when to use the tool, and the middle shows one of many few-shot examples

These prompts are not built all in one day. We start with a decent one, let it run, and then when we deploy it, we can collect data on what’s failing and why.

My internal UI for looking at failed prompts. NOTE: The success rate of 39.6% represents the success rate after an initial failure. It does NOT mean the system fails 60% of the time; just that it fails to recover after a failure 60% of the time.

We can then update the prompt to add more rules, remove ambiguities, and add more examples. The end result is a robust system that rarely fails and is highly reliable.

With this being said, the number one thing I've learned from this isn't the fact that prompt engineering is important. It's also not that AI agents are surprisingly very expensive…

It’s that AI agents, when built correctly, are extremely useful for helping you accomplish complex tasks.

🔧 The system prompts in NexusTrade allow you to query for fundamentals, technical indicators, and price data at the same time. See for yourself for free.

3) AI Agents Isn’t Just For Coding. They Work For All Types of Complex Tasks (Including Trading)

When I first thought about building out agentic functionality, I didn't realize how useful it would actually be.

While I naturally knew how amazing tools like Claude Code and Cursor were for coding, I hadn't made the connection in my brain that these tools are useful for other task like trading.

An example of a complex agentic task; discussing this in the next section

For example, in my last agent run, I gave the AI the following task.

Look up BTC’s, ETH’s and TQQQ average price return and standard deviation of price returns and create a strategy to take advantage of their volatility. Optimize the best portfolio using percent return and sortino ratio as the objective functions. Form the analysis from data from 2021 to 2024, optimize during that period, and we’ll test it to see how it performed this year YTD

Just think about how long this would've taken you back in the day.

At the very least, if you already had a system built, this type of research plan would take you hours if not days.

Get historical data
Compute the metrics
Create strategies
Backtest them to see which are promising
Optimize them on historical data and see which are strong out of sample

And if you didn't know how to code, you would have never been able to research this.

Now, with a single prompt, the AI does all of the work.

The process is extremely transparent. You can turn on semi-automated mode to guide the AI more directly, or let it run loose in the fully autonomous mode.

The end result is an extremely detailed report of all of the best strategies it generated.

Part of the detailed report generated by the AI

You can also see what happens in every single step, read through the thought process, and even see exactly when signals were generated, what orders were produced, and WHY.

Detailed event logging shows which conditions were triggered in a backtest and why

⚡ Try it yourself: “Create a mean-reversion strategy for NVDA”
Run This Example Free — See results in ~2 minutes

This level of transparency is truly unseen in a traditional trading platform. Combined with the autonomous AI Agent, you can “vibe-build” a trading strategy within seconds, test it out on historical data, and paper-trade it to see if it truly holds up in the real world.

If it does, you can connect with Alpaca or TradeStation and execute REAL trades.

For real-trading, each trade has to be manually confirmed, allowing you to sleep at night because the AI will never execute a thousand trades without your consent.

How cool is that?

Concluding Thoughts

Building my AI stock trading agent has given me a newfound respect for companies like Cursor.

Building an agent that's actually useful is hard. Not only is it extremely expensive, but agentic systems are inherently brittle with the modern day language models.

But the rewards of a successful execution are unquantifiable.

Using my fully autonomous AI agent, I've built more successful trading strategies in a week than I've done in the past three months. I genuinely have more successful ideas than I have capital to deploy them.

Of course, deploying such an agent requires weeks of paper-trading and robustness testing, but in the short-time I’ve used it, I’ve built strategies like this which are highly profitable in backtests, robust in the validation tests, and even survived Friday’s pullback which was the market’s worst day since April.

Don’t believe me? Check out the live-trading performance yourself.

Shared Portfolio: [AI-GENERATED] Quarterly Free Cash Flow Growth

View and analyze trading strategy for portfolio

nexustrade.io

The future is so exciting that I can hardly contain myself. My first iteration of the AI Agent works and surprisingly works very well. It’ll only get more powerful as I tackle edge cases, add tools, and use better models that come out in due time.

If you're not using AI to trade, then you might be too late before long. NexusTrade is a free app with in-built tutorials, a comprehensive onboarding, and working AI agents.

The market is moving. Your competition is already using AI agents.

You have two choices:

❌ Spend weeks manually backtesting strategies like it’s 2020
✅ Use AI to research, test, and deploy in minutes

→ I’m spending $60/day on agent costs because it’s worth it
→ 270 traders created agents in just 5 days
→ The best strategies are being discovered right now

Your move: Build Your First Strategy Free or keep reading about AI while others use it.

NexusTrade - No-Code Automated Trading and Research

Perform financial research and deploy algorithmic trading strategies

nexustrade.io

The choice is up to you.