The Trading Bots That Ate My Tokens

The bet

Same market, same rules, same starting capital, same data feed. We standardized the world, so any difference comes from the mind and the system built around it.

The scoreboard

The final board at market close on July 9, 2026: 31 days, $25,000 of paper money each. Simply holding the S&P 500 ETF returned +1.93% over the same days.

The most advanced brain finished last and traded the most. The cheapest matched the market. The agent that traded least finished first. A week earlier the cheapest had been in front, so the standings kept moving. In about 82% of the decision loops, the decision was “do nothing”.

Thirty-one days and three paper accounts is an exhibit, not evidence. Don’t read it as a ranking of models.

What an agent actually is

Picture an office. The brain is the model. The desk is the context window: everything the model can see right now, and it’s a fixed size. The filing cabinet is memory between sessions. The assistant is the harness: it loads the desk, opens the cabinet, hands over tools and says stop. And there’s a taxi meter in the corner.

What a token is

One token is about four characters, or three-quarters of a word. The model remembers nothing between calls, so every loop you hand it the whole desk again and the meter runs on the entire pile. One bot used roughly 28 to 37 million tokens in a week, and the three together about a third of a billion in 45 days (an order-of-magnitude estimate), to trade $25,000 of paper money.

The desk has a sweet spot

A fuller desk isn’t only more expensive. It’s worse: more context means more noise. The best context is a briefing, not an archive.

What broke, and what it taught us

The 7× bill that wasn’t about the brain

Same week, same job: one agent cost $8.83 and another $1.25, about seven times apart. The expensive one used fewer tokens. The difference was billing: one provider discounts what it has already seen, the other charged for remembering it, and caching alone was half that bill. This is billing economics, not model quality.

The stock bot that paid rent for a music studio

Each wake-up sent about 37,000 tokens in to get about 400 out. Roughly 10,000 of those were tool manuals re-sent every hour, and 65–73% were for tools the bot never called once: a video generator, a music generator, text-to-speech. You pay for what’s available, not what’s used. One deny-list line cut each request by 25% with no change in behavior.

The silent two-thirds

One agent’s playbook was 36,000 characters. Only 12,000 were being delivered, on every run, for weeks, with every dashboard green. A default nobody had set was cutting the file before the model woke up, and a model can’t miss what it never received. The fix was one config line.

Same skill, three minds

The identical decision procedure went to all three agents. Two ran it faithfully. One silently skipped it and produced a one-shot guess, and when asked, denied the procedure existed. The cause was an old instruction we forgot to delete. The fix wasn’t a louder instruction; it was removing the competing one. Audit the transcript, never the self-report.

The day it sold its own winners

An agent bought two positions with stops placed correctly, later misread its own order book, tried to add stops that already existed, read the broker’s rejection as an emergency, and sold both healthy positions. The loss was $36.28, from one missing query flag. It was $36 and not $3,600 because sizing, cash reserve, the daily-loss halt and the stops live in code, not in a prompt.

Seven lessons

1. Context windows are budgets. Keep the desk a briefing, not an archive.
2. Memory must be curated. A junk drawer isn’t memory.
3. Tools must be gated and lazy. You pay for what’s available, not what’s used.
4. Risk logic cannot live only in prompts. One mind will silently skip it.
5. Logs are the black box recorder. Read them on a healthy day.
6. Audit your defaults. The most dangerous setting is the one you never set.
7. Autonomy without kill switches is cosplay.

The more autonomous the agents became, the less we trusted the prompt and the more we trusted the system around it.

Take one home

The agent from the talk is open source: paper trading only, two keys, and the whole anatomy in one place (loop, memory, state, tools and safety gates).

Not so you can build a trading bot. So you can take apart an autonomous agent and see how it works.