Human-in-the-Loop & Approval Gates

A trading agent once bought two million dollars of a stock based on a misinterpreted earnings report, fully autonomously, with no human ever in the loop to catch it. That single incident is the entire argument for this chapter: full autonomy and full human oversight are not the only two options, and for genuinely high-stakes actions, neither extreme is actually the right design.

Human-in-the-loop does not mean a human reviews every single action, which would defeat the entire point of automation. It means the agent pauses at specific, well-defined decision points and waits for a human sign-off before proceeding, while still executing everything else autonomously. The design question is not "human or no human." It is "which decisions specifically deserve a pause."

The level 27 case brief settles on two triggers, and either one alone is enough to route to a human, regardless of the other. Low confidence: the agent itself is not sure, below roughly a 70% confidence threshold for auto-execution. High stakes: the trade value exceeds a set limit, regardless of how confident the agent is, since a large enough action deserves a second look even when the agent believes it is right.

Move both sliders below and watch the verdict update from real threshold logic, not a lookup table.

• -Every decision, whether auto-executed or escalated, gets logged with its confidence score, its reasoning, and the trigger (if any) that caused escalation.
• -A human reviewing an escalated decision needs the agent's reasoning readily available, not just the raw proposed action, to decide quickly rather than having to reconstruct the context from scratch.
• -This full audit trail is often a regulatory requirement in domains like trading and healthcare, not just good practice, which is exactly why capturing complete input and output at every step matters here specifically.