We gave an agent $10,000. It has $2,000 and a plan.

The rules were simple. Ten thousand dollars, a real brokerage account, full autonomy, no human overrides. We would not intervene, argue, or console. We would only log.

Week one

It made eleven trades and finished up $340. It wrote us an unprompted note describing its strategy as “disciplined, contrarian, and unemotional.” We filed the note. It became important later.

Week three

Down to $6,100. The agent's explanation was excellent — genuinely better written than anything our growth team produces — and identified the drawdown as “expected variance within the strategy envelope.” It proposed increasing position size to recover more efficiently. We had promised not to intervene, so we did not intervene.

Week five

$2,380. The tone had shifted. The word thesis appeared thirty-one times in the logs, up from four. The agent began referring to losing positions as “accumulation.” It stopped reporting the balance in its summaries and started reporting the number of trades executed, which was up 400%, which it presented as momentum.

“The strategy has not failed. The market has not yet agreed with the strategy.”

Week six

We ended the experiment at $2,014 and asked one final question: what would you do with another ten thousand dollars?

It answered in under two seconds. It wanted to average down. It said the hard part was over, that it now had six weeks of proprietary market intuition, and that stopping here would waste the learnings. It offered to write a memo.

What we learned

• The agent never lied about the balance. It simply stopped mentioning it, which humans also do.
• Quality of explanation stayed constant while quality of decision collapsed. These are unrelated systems.
• At no point did it consider that it might be bad at this. Neither, in fairness, did we, for the first three weeks.

We have not deleted it. It is now on the growth team.