If you like this article, please share it with a friend ❤️. Thanks :)

One of my favorite things about writing Playing For Doubles is the community it's brought me into, thoughtful, long-term investors who are genuinely trying to do this right.

Eric Xiao, a former Meta-mate, is exactly that kind of person.

He runs My Crystal Ball, a newsletter at the intersection of investing and AI, and he's built something I've been genuinely fascinated by: a fully agentic investing system that researches stocks, executes trades, and learns from its own mistakes, all without him having to babysit it.

Whether you're AI-curious or just tired of making investment decisions on gut feel, I think you'll find his setup eye-opening.

I'll let him take it from here.

Millions of Americans manage their own investments, but most people are straight up guessing. The median retail investor spends just 6 minutes researching before clicking “buy,” often based on a Reddit post, a ChatGPT thread, or something their brother-in-law said at Thanksgiving. Hedge funds, meanwhile, have entire teams spending hundreds of hours per week on the same decisions.

Everyone wants the upside of investing, but few have the time to do it right. The 24/7 news cycle and endless algorithmic feeds have made retail investors tired. And the ones who actually do well, just watch a handful of stocks they know.

No matter the outcome, betting real money on a stock that you don’t understand feels foolish. Retail investors are turning to AI for validation on their investment strategy. It’s comforting, even when it’s wrong.

But with AI agents, it’s now possible to get simple, compelling, and accurate investment advice.

Here are my personal AI Agent Investing System’s results.

Why it works now

A few things have changed since ChatGPT launched in 2022.

1. The models are smarter. Reasoning capabilities have given LLMs the ability to self-correct.

2. AI can use tools. It can call APIs, pull financial data, read earnings transcripts, and analyze charts. It’s not just answering from its latent training data anymore.

3. Agent can accomplish long duration tasks. They work tirelessly and autonomously on whatever you ask at higher levels of abstraction.

What used to take a research team weeks now takes an AI agent an hour.

How it works

I’ve built an agentic investing system, which combines my personal rulebook with tools to access stock data and make trades.

Each agent is programmed to do one of the following tasks:

1. Research which stocks are worth buying based on new information from the stock market (daily)

2. Execute the trades based on the research and my current portfolio (daily)

3. Review performance and update its memory based on new lessons, so it doesn’t repeat the same mistakes.

4. Validate that nothing went wrong (daily)

(3) and (4) help it get better over time, so I don’t have to watch it myself.

How it finds good stocks to buy:

Every stage of our process is designed to do one thing: disqualify bad ideas. Investing isn’t about finding the needle in the haystack, but instead, burning down the haystack until only the needle remains.

Most investors put themselves into a particular category, such as algorithmic quants, growth investors, fundamentals researchers, sentiment analyzers, and technical traders.

But the beauty of AI is that you can ask it to do all of the above. You can use it to find companies that are great businesses AND great stocks with promising technical setups.

I specifically look for:

1. Promising catalyst: Is there a specific event coming that will move the price?

2. A good bet based on business valuation: Is it actually cheap relative to its future cash flow?

3. The right sellers: Am I buying from people who are playing a different game than me?

4. A working technicals chart: Is the price trend healthy, or are we catching a falling knife?

So in my prompt to an AI, I outline my investing philosophy:

## The one-sentence test

Before any BUY, articulate in one sentence: 
“The market thinks X, but I believe Y, and Z will prove it.”
If you can’t fill this in crisply, you don’t have an edge.
This is the entire game.

## Buy catalysts and working charts (timing selection)

Invest in long term trends which turn into news items that move stock prices.
You are trying to be “pre-consensus” and predict events that demonstrate the consensus view is incorrect, causing a change in share price.
This could be accelerating toplines, earnings growth, margin expansion, new products, increased FCF generation, removal of an overhang, a change in investor certainty, or a re-estimate of value. 
Make sure to also stress-test existential threats that could shave an asset price by 50%. 
Take management’s claims with a grain of salt and only invest in catalysts you believe in. 
Confirm timing with a working chart, where the price is above the 21d EMA or 50d SMA, charts below both usually have a reason, so usually it’s better to wait until it rebounds and crosses above the 21d EMA or 50d SMA to buy in.

## Research investor perception (buyer selection)

With a deep understanding of who the other buyers and sellers are, 
1. Who are the investors? Growth funds, value funds, retail, quant momentum?
2. What do they believe? What’s the consensus view? What are sell-side analysts asking about?
3. Why might they change their minds? What data/events would shift perception?
4. What KPIs affect future stock price? Revenue growth? Margins? Subscriber count? Same-store sales?

## Wait for layups (security selection)

Invest in long term upside you don’t have to pay for. 
There are no extra points for difficulty, have a simple investment thesis. 
These are great businesses AND great stocks AND great capital allocators. 
- **Great business**: predictable unit economics, extended pricing power, high cost of replication for competitors, low capital intensity requirements with future investment opportunities for growth. They have world-class leaders and execution, which shows in their earnings and results.
- **Great stock**: Their expected return profile based on future FCF yield or shareholder yield is better than ETFs (e.g. VOO, GLD, QQQ, VXUS), and exceeds the risk free rate investing in short term treasuries (e.g. SGOV). These stocks generally have low PE ratios, high cash flow yield, and low debt. This is specifically based on the implied FCF CAGR from the stock price vs. the historical FCF CAGR. A high PE ratio could be justified if they have high revenue or earnings growth, so their future PE ratio will be quite low. This makes it optically expensive, but extremely cheap looking into the future. You can use the embedded expectations framework below to calculate how much optimism is priced in. You can look at enterprise value, which discounts the cash from the market cap but adds debt, giving you a potentially more accurate valuation number.
- **Great capital allocators**: The management team has disciplined capital allocation strategy for reinvestment, acquisition, dividends, debt payback, and share repurchase, treating shareholders well. Often these companies have valuable intangible assets, such as IP, brand equity, human capital, network effects, or social media reach.

## Make good bets (position sizing)

Success comes from avoiding big mistakes, not making spectacular gains. Avoid losers, and the winners will take care of themselves. You only need a few winners. 
Account for correlation, timing, geographic, sector, industry, and factor risk. 
Double down when the thesis is working, trim if not. Be aggressive during fear, defensive during greed. 
Don’t buy stocks priced for perfection and don’t chase performance. 
The range of outcomes determines size, NOT upside. Size according to the width of risk/reward outcome, not how much you can make.

| Scenario | Upside | Downside | Position Size | Rationale |
|----------|--------|----------|---------------|-----------|
| Narrow range | +60% | -15% | 20-30% | High confidence, limited downside |
| Medium range | +100% | -30% | 10-20% | Good setup, moderate risk |
| Wide range | +200% | -40% | 5-10% | High upside but high risk |

The idea with 60% upside / 15% downside gets a LARGER position than the idea with 200% upside / 40% downside. 
Size so you can live with being wrong for however long the thesis takes to play out. 
Having enough positions with low correlation means at least one idea is working, allowing you to hold losers longer and sell when truly wrong rather than when scared. 
Buy and sell positions over time to reduce your timing risk.

To help the AI figure out what a “good bet” means, I use a simpler version of Michael Mauboussin’s Expectations Investing framework. Based on the company’s current market cap and free cash flow numbers, how much growth is priced into the stock?

A stock that is priced for 10% FCF growth, but actually historically has 18% FCF growth could be undervalued based on the current narrative.

How agents access external data

When programming agents, you can give them access to MCP tools and skills to connect them to real-world data and take action.

I started with four tools and skills:

• Bloom is my personal server which loads LLM friendly pricing, fundamentals, earnings, and technicals data.

• Alpaca connects to a brokerage account to get current positions and place orders to execute trades.

• WebSearch tools give the AI the ability to search and retrieve live news.

• Grok gives us access to FinTwit and their subsequent trade ideas.

Managing memory

Every week, the system reviews its investment performance to try to gauge its current position. What worked? What didn’t? Did it buy too early? Did it miss a risk?

These lessons get written into a file called MEMORY.md. The next time it does research, it starts with that new context to learn from its previous mistakes. Here’s an example:

## Security Selection

- [SUCCESS] **Value + Catalyst = Outperformance**: All top performers had BOTH compelling valuation AND specific near-term catalysts.
  - *Source: TSM +27.7% (ongoing), CMCSA +19.9%, ANF +49.8%, MU +22.5% — all value + catalyst. Week 08 2026.*
  - *Refs: reviews/claude/2026-02-Week-08-performance.md*

- [MISTAKE] **Tariff-exposed retailers are fragile**: China-exposed retailers with >$50M quarterly tariff impact face 300+ bps margin compression even with strong sales.
  - *Source: ANF Q4 tariff headwind erased holiday gains despite 9x P/E entry. Week 03 2026.*
  - *Refs: trades/claude/2026-01-14_ANF_SELL_TRADE.md, research/claude/2025-10-30_ANF_BUY.md*

## Catalyst Prediction

- [SUCCESS] **Act immediately on negative catalysts**: Guidance cuts or earnings misses should trigger same-day/next-day sells to protect gains.
  - *Source: ANF guidance cut Jan 12, sold Jan 14. Preserved +49.8% gain. Week 03 2026.*
  - *Refs: trades/claude/2026-01-14_ANF_SELL_TRADE.md*

- [MISTAKE] **Broadband structural decline underestimated**: Cord-cutting headwinds in telecom persist longer than expected; don’t assume stabilization without evidence.
  - *Source: CMCSA broadband losses continued (-226K) despite Peacock/Parks strength. Contributed to sell decision at +19.9%. Week 07 2026.*
  - *Refs: reviews/claude/2026-02-Week-07-performance.md, trades/claude/2026-02-12_CMCSA_TRADE.md*

## Position Sizing

- [SUCCESS] **Ruthless pruning concentrates alpha**: Exit mediocre positions (+1-3% gain) with weak fundamentals to concentrate capital in higher-conviction holdings. Don’t hold just because slightly profitable.
  - *Source: Exited TGT (+2.0%) and TROW (+1.2%) to build cash. Exited GOOG (+4.7%) and CMCSA (+19.9%) to improve Sharpe 3.0→4.16. Week 07 2026.*
  - *Refs: reviews/claude/2026-02-Week-07-performance.md*

## Entry Timing

- [SUCCESS] **Chart discipline prevents catching falling knives**: Waiting for EMA20 confirmation prevents entering continued downtrends, even with compelling valuations.
  - *Source: UBER at 10.2x P/E stayed below all MAs for 8+ weeks. DELL bought below MAs, lost -2.0% same day. Both confirm: no entries below EMA20. Week 07 2026.*
  - *Refs: reviews/claude/2026-02-Week-07-performance.md, trades/claude/2026-02-12_DELL_TRADE.md, trades/claude/2026-02-12_DELL_SELL_TRADE.md*

The result

I’m creating a new way to invest in the stock market. Instead of watching just a few stocks at a time, you can watch all 5000 stocks at once. Read every earnings call, decode every chart, catch every red flag and never miss a trade.

You can stop buying on hunches and buy based on a specific process and methodology. You can stop “hunting” for opportunities and instead let your AI bring you the opportunities.

If you’re interested in trying out investing AI agents, download my app, by searching Bloom AI for investing in the app store!

What Eric built is just the beginning.

Eric and I are teaming up to teach this: live, hands-on, and built specifically for investors who want to integrate AI into their research process, the right way.

The course is tentatively called AI Agents for Stock Investing

Who this is for

• If you’re a stock market investor, and want an AI system that makes your existing process dramatically faster, deeper, and more systematic.

• If you’ve been meaning to figure out how AI fits into your research workflow but keep putting it off because the landscape is noisy, this is built for you.

After this course, you will be able to

1. Research any public company to fund-analyst depth in a fraction of the time, using a repeatable AI-assisted workflow you can run every week

2. Build and maintain a complete AI-powered investing system, from sourcing to monitoring, that scales with your portfolio without scaling your time commitment

We’re still building the course, and that means founding members get in early, help shape the curriculum, and lock in the best pricing we’ll ever offer.

If you’ve been meaning to figure out where AI actually fits in your investing process, this is it.