one hour                person        agent
   actions                    12          1400
   distinct resources          5           260
   gap between calls        180s             2s
   sessions                    1            96

   chapter 8   detect it     chapter 9   respond to it
   +----------------------+  +-------------------------+
   | triage as a loop     |  | scope at machine speed  |
   | detections FOR agents|  | contain, replay, stop   |
   +----------------------+  +-------------------------+

CyberTravels has a SOC. It was built for people.

It watches for a login from an unusual country, a burst of failed authentications, an employee downloading the customer list on their last day. It is good at those, and none of them describes CyberTravels.

An agentic SOC watches a different actor. One hour of CyberTravels' Workflow Agent is roughly 1,400 tool calls across 260 resources in 96 sessions. One hour of Alex is twelve actions. Every threshold, baseline and playbook CyberTravels owns was tuned against the second number.

And the actor is not only the adversary. It is also the instrument: chapter E2 puts an agent on detection engineering and chapter E3 puts one on the alert queue. Both work, and both bring the same failure mode — a loop that concludes confidently can conclude wrongly at machine speed.

The unit this function is written in

A ratio of 117× is not a fact you can act on. The unit is an interval, and there are five of them between an agent doing something it should not and the control that stopped it being back at target:

Every lesson in this function shortens one of those five, or spends one deliberately to buy something else — and each says which, in a line under its own concept. The intervals are where the numbers live: CyberTravels' detection interval is 194 minutes against a 15-minute target, and its manual containment runbook finishes in 34 minutes against a measured breakout time of 29. Neither of those is an opinion about tooling.

The stack that runs it

Nothing in this function needs a product you do not have. Every phase below has a working open-source reference, and the lessons are written against these rather than against a vendor's diagram — so you can build the whole thing and find out what it does not cover, which is the part that matters.

Two honest notes about that table.

The DLP row is the weakest. There is no open-source DLP with the maturity of the other three, which is why the row names a pattern rather than a product — and why E1.1 scores DLP at almost no coverage of an agent's day. That is a real finding about the market, not a gap in the reading.

The agent-telemetry row does not exist in any of the products. Wazuh will tell you a process wrote a file. Nothing in the list tells you which prompt caused it. That row is instrumentation you write, it lands in the lake E2.1 designs, and its absence is what E1.1 measures.

The five chapters

E1 — discover. The four sensor classes you already own, scored against what an agent actually does; drift, the failure with no adversary at all; and a bonus on finding the agents nobody registered and keeping what they emit.

E2 — detect. The lake every rule is written against, then detections for two subjects that are not the same subject — the agent, and the platform running it — mapped to ATT&CK and ATLAS, plus the loop that writes rules, the benign corpus that decides whether they ship, and the one detector that needs no threshold.

E3 — understand. An investigation an agent can run: bounded before it starts, given the fields an agent alert needs, willing to abandon its first theory in the open, scoped along the delegation graph, widened to the population — and then the two proactive halves, third-party intelligence that has to become a rule and a hunt for behaviour no rule covers.

E4 — respond. A response whose blast radius is known before it fires. Actions classified on reversibility and radius, tiers derived from that rather than from their author, containment timed against the attacker, and a fleet stop that revokes as well as terminates.

E5 — recover and root cause. A run you can reproduce, a root cause that names a control rather than a person, the fix put at the layer it belongs in, the indicators re-measured to see which actually came back — and a regulatory clock that started before anyone knew.

Two rows of the CyberTravels register are this function's whole reason to exist: R9, where holding one agent reaches CRM, payroll and the cloud resource manager, and R10, where passport numbers are in the logs that would have told you.