How TradePass runs
We build TradePass to operate itself: an AI agent fleet handles content quality, tutoring, support, and operations. This page is our honest scorecard.
Human interventions this month
10
Times an agent stopped and a person had to step in, because it couldn't safely act on its own.
Ad-account actions taken autonomously this month
1
Negative keywords, bid changes, and new keywords our ads agent applied directly to the live account — never touching daily spend limits, always logged with a reason.
What we are not counting above, and why
An honest counter is only honest if it says what it excludes. Two other things get logged to the same audit trail this month, and neither is a person stepping in:
- 30 advisory handoffs — the ads agent finished its daily review and left a recommendation on a decision it is not permitted to make itself, such as changing a daily spend limit. A handoff, but a routine one, not a failure.
- 62 automated rejections — unauthenticated requests to our support endpoint, refused by code. Nobody saw them. They are a security statistic, not an operations one.
Until 2026-08-12 this page added all three together and called the total “human interventions”. That overstated how much we depend on people, so we split them out rather than quietly shrinking the number. Any new kind of handoff we start logging counts as a human intervention by default until we deliberately classify it otherwise.
What the agents do
- Question factory — generates practice questions, independently re-verifies every one with a second model, and re-executes every calculation in plain Python (never trusted to the LLM).
- Tutor — answers "why" for any question, grounded to the verified answer, scoped to that question only.
- Support — reads inbound email, answers product questions, and issues refunds automatically when a request clearly matches our published policy. Anything ambiguous is escalated to a human — counted above.
- Ops watchdog — reviews production logs daily; genuine problems are escalated, not silently retried.
- Ads agent — reviews Google Ads keyword performance daily on a 5-day rolling window, applies negative keywords, bid adjustments, and new keyword additions within fixed guardrails (minimum sample size, capped change size, daily action limits, cooldowns). Daily spend limits are never touched by this agent — that stays a human decision.
We'd rather escalate too often than get an edge case wrong — the counter above is not something we're trying to shrink to zero at the cost of good judgment.