Detection Latency

By the time an incident is triaged, the agent has finished its task. The median human response to an alert is measured in minutes at best. An agent loop completes in seconds. Detection that lives outside the loop is not early warning. It is archaeology.

The math of the gap

The numbers from the field make the gap concrete. Institutions report that agents already call call centers to move money. Attacks arrive with zero simmer time: explode, impact, disappear. The window between action and evidence was always the design constraint of detection. The window is gone.

A human fraudster cased the target, built the tooling, moved slowly enough to leave a trail worth analyzing. The agent loop compresses all of it. By the time a human reads the alert, the task is done, the funds are moved, and the loop that produced the alert has terminated cleanly.

Why outside-the-loop detection fails

Traditional detection assumes the investigation starts while the attack is still unfolding. That assumption carried decades of security operations. It does not survive machine speed, because machine speed removes the during. There is only before and after.

Post-hoc detection still has value: forensics, pattern extraction, hardening the next loop. But it cannot interrupt, and interruption is the only response that changes the outcome for the deployment that was hit.

What inside-the-loop means

Instrumenting the loop itself: the tool calls, the thresholds, the authorization checks. The agent's own sequence of actions is the highest-fidelity signal available, and it is only available in real time.

The engineering cost is real. Latency budgets, false positive tuning, threshold ranking. But the alternative is a detection stack that reconstructs incidents beautifully and stops none of them. Detection inside the loop turns incident response from reconstruction into interruption. That is the difference between a postmortem and a non-event.