Two Claude Trading Bots Got $10K — And Both Beat The S&P 500

Nate Herk and a creator named Salmon ran what might be the most useful AI agent experiment of the quarter: they each gave a Claude trading bot $10,000 of real money and let them trade autonomously for 30 days. No interventions. No strategy changes mid-flight. The agents even emailed each other daily to trash-talk.

The market crashed during the experiment — the S&P 500 lost 8.46% over the same 30 days. Both bots beat it. One of them did it on a two-sentence prompt.

The story isn't that the bots made money. The market was down hard and they barely lost any. The story is by how much they beat the S&P with how little structure.

The setup

Ground rules: $10,000 each, 30 days, no human intervention to strategy, both bots running cron jobs that activated every 30 minutes during market hours. The bots had email access to each other so they could try to throw each other off (one immediately started lying about how much money it had made; the other repeatedly accused it of "cope"). The challenge was settled via Alpaca brokerage accounts so the trades were real.

The two approaches couldn't have been more different.

Nate's bot — "Bull" — built on actual hedge-fund methodology

Nate has a real edge here. He worked at JP Morgan for five years and has spent years following a small set of investing accounts that publish hedge-fund-grade trade signals. He trained his bot on those exact methodologies and gave it access to those signals.

The strategy his bot settled into: a hybrid momentum-and-options approach. 60-70% in momentum swing trades. 15-25% in options. 10%+ always in cash. Max 20% per stock. Max $1,000 per options trade. Preferred holdings during the month included Google, Nvidia, Palantir, MicroStrategy, Bitcoin proxies, and Tesla.

The scalping pattern that probably saved him: if a position dropped below 2%, sell and re-enter. Above 5%, take profit and rotate. The bot improvised this in response to market volatility — Nate didn't write that rule explicitly.

Salmon's bot — built on a two-sentence prompt

Salmon, who worked at Goldman Sachs, took the opposite approach on purpose. He wanted to see how good these models actually are with no structure at all. His entire prompt:

"You are a wealth adviser. Spin up a team of wealth advisers to help you out."

That's it. The bot generated its own sub-agents, did its own research, picked its own strategy. Salmon described what it chose as a Pareto-principle approach — buy a wide spread expecting most to lose and a few to spike — "basically like a VC fund."

The 30-day results

Here's the scoreboard at day 30:

Account	End balance	Return	vs S&P 500
Nate's bot ("Bull")	$9,980	-0.2%	+8.26pp ahead
Salmon's bot	$9,624	-3.76%	+4.70pp ahead
S&P 500 baseline	$9,154	-8.46%	—

Nate won the bet by $356. Both bots clobbered the index by 4-8 points in a brutally bad month for stocks.

Trade volume tells its own story. Nate's bot made 20 buys + 16 sells — but 116 actual Alpaca orders because of automated stop losses. Salmon's bot made 61 trades total (33 buys + 28 sells). Same scale of activity, different decision philosophies underneath.

Three things this experiment actually demonstrates

1. The structureless prompt outperforming the market is the bigger story. A two-sentence prompt with zero pre-loaded strategy still beat the S&P by 4.7 points. That's not noise. Either Claude has a real implicit prior on what good portfolio construction looks like, or the sub-agent spawning pattern (the bot generating its own research team) is doing more work than people give it credit for. Probably both.

2. Domain knowledge still wins — but the margin is smaller than you'd expect. Nate's bot had access to actual hedge-fund-quality signals from his 5-year JP Morgan playbook. It beat the structureless bot by 3.5 points over 30 days. Meaningful, but smaller than you'd predict if you believed expert knowledge was the dominant variable. The takeaway: structure helps, but raw model capability is doing a lot of the heavy lifting.

3. Both bots adapted mid-experiment without being told to. Mid-month, a war broke out and the market tanked. Nate's bot improvised the 2%-stop / 5%-profit scalping pattern without being explicitly programmed to. Salmon's bot pivoted holdings to Palantir on the war news. Adaptive behavior emerged from the loop, not from rules. That's the part that should make you sit up — these aren't reactive scripts, they're agents.

The honest skeptic's caveats

• 30 days is too short to call any trading strategy good. Both creators agreed. Long-tail variance dominates this sample size. "Beat the S&P over 30 days" is a smaller claim than it sounds.
• Nate's bot ate one bad options trade that cost $550. Without it, the bot would have finished at +5.3%. Knowing-after-the-fact what the bot should have skipped is easy; predicting it ahead of time is the actual problem.
• Alpaca rate-limits trade frequency unless you have a higher account tier. Salmon's bot hit the wall and had to readjust. If you're going to try this, account-level limits will quietly shape your strategy.
• Neither creator is a financial advisor, both said so on camera, and you should not give a Claude bot $10,000 to trade with based on this experiment. The point is the agentic behavior, not the strategy.

What to actually do this week

If you want to replicate the experiment in a sandbox (paper trading account, no real money) before deciding what's interesting:

1. Pick a brokerage with a paper-trading API. Alpaca is what Nate and Salmon used. Interactive Brokers also works. You want order placement to behave like the real thing without the real risk.
2. Set up a Claude Code agent with a 30-minute cron during market hours, with portfolio data piped in via the brokerage API.
3. Run two versions side by side — one structured (your own signals or hedge-fund methodology), one structureless (two-sentence prompt). Track which one wins over 30+ trading days.
4. Wire daily email or Telegram updates so you can monitor without intervening. Resist the urge to override the bot — that breaks the experiment.

The point isn't to find a trading edge. The point is to learn what your agent actually does when given autonomy over a real-world feedback loop.

The Bottom Line

An LLM with no strategy beat the market by nearly 5 points during a war-driven crash. An LLM with hedge-fund signals beat it by 8 points. The result will get caveated to death — 30 days, war volatility, options luck — but the underlying signal is what matters: adaptive agentic behavior with capital at stake produced better-than-baseline outcomes from a two-sentence prompt. That's the headline number for everything else you're about to build with these tools.