The First Bot (Grid)

Two complete round trips on $500 realize about seventeen cents. Each of the four trades behind them costs a conversation with a language model, drawn from a budget that refills every five hours whether or not the strategy is working.

The grid bot did exactly what I built it to do. That turned out to be the problem, and it is why the bot was retired after a week.

Why a grid, and why TQQQ

• Capital. Start small. Paper trading is great to learn the fundamentals, but I wouldn’t have skin in the game
• Frequency. Need some volatility. Something that would trigger buys/sells a couple of times a week.
• Parameters. Buy as price dips, sell small chunks as price rises

I looked into leveraged Exchange-Traded Funds (ETFs). TQQQ is 3x daily QQQ. Let’s experiment with $500.

The grid bot would execute on a tight range: sell 3% of the position on any rise of 0.05% or more, and re-buy with the sale proceeds on any fall of 0.05% or more. The threshold is a floor, not a target: a move straight from $70 to $68 clears it just as a 0.05% tick does, so trades can fire at much larger swings than the trigger implies.

What the grid actually earns

Two round trips, priced out end to end. What does a trade at this size actually put in the account?

Illustrative walkthrough of the grid logic, not a log of executed trades. The larger moves still clear the 0.05% floor.

The mechanism works. Each trim sells into strength and buys the dip back with the same dollars, so the share count creeps up: 7.152 shares at the end against the 7.143 bought on day one.

The bot was running on its own. No need to monitor its activities. So far so good.

What I Had Not Accounted For

The two things that killed it were token usage and volatility decay.

Token burn

Trading with such frequency burned through Claude Pro’s token budget. Each trade requires a conversation with AI to execute. AI checks the price, evaluates the strategy, places the order, confirms execution, then updates its records. The grid bot could trigger several times an hour, quickly reaching token usage limits within Claude’s 5-hour window.

This is the constraint that has no analogue in traditional algo trading. A conventional bot’s cost per trade is commission and slippage. An agentic bot pays in tokens, and that budget is fixed per window regardless of how well the strategy is doing. Trade frequency stops being purely a strategy decision and becomes a budget decision.

Volatility decay

TQQQ can move +/-1% in a sideways market. The underlying QQQ ETF could go up 1% then down 1% over two days which would be back to roughly even. Since TQQQ is 3x the movement, the same two days is +3% then -3%, which does not return to even. It nets about a 0.10% loss. Repeat that for a month and the gap compounds.

If the index goes nowhere, where does a 3x fund end up?

QQQTQQQ (3x daily)

Illustrative: a perfectly alternating market, indexed to 100. Real chop is messier and the drift is not this tidy.

Twenty flat days cost QQQ 0.1% and TQQQ 0.9%. A grid bot that buys every dip in that market is buying an asset whose floor is sinking underneath it.

After a week of analyzing the grid bot’s activities, the grid bot was retired and converted to a B&H strategy.

---

Next: if frequency is expensive, the trades have to be worth more. The next bot moved to SOXL, dropped the trigger from 0.05% to a 7% daily drop, and sold in bands instead of fixed 3% trims. That is Dip Accumulators.