Agentic Security
Engineering.
The field, defined — with its craft, its laws, and its anatomy.
Agentic security engineering is the discipline of integrating an agentic layer from first principles, as a first-class citizen of any defensive security system.
The term is often read narrower than it is. Two nearby ideas keep getting attached to it:
Securing the models themselves — hardening against prompt injection, jailbreaks, and the like.
Defending against AI-driven attacks — the adversary pointing models at you.
The field includes both — a system that reasons with models carries them in its attack surface, and AI-augmented adversaries are table stakes. But it is more than the sum of those parts: the heart of the discipline is the defensive system itself, with the agentic layer working inside it.
If agentic security engineering is the what, applied harness engineering is the how.
Applied harness engineering is a way of thinking: knowing when, where, and how to integrate an agentic layer into a defensive system — and, just as importantly, when not to.
There is no one correct architecture to hand over — the right answer depends on your telemetry, your threats, your constraints. The craft is the way of thinking you apply to your own context.
The model you select — you will realistically never train a frontier model, and beyond selection you change little. The harness you engineer — it is just an application, built exactly how you want.
A harness is the deliberate orchestration of three collaborators — human judgment, agentic reasoning, and deterministic code — each used only where it is strongest, and designed that way from the ground up.
Collaboration is first-class, by design
The human–agent relationship is part of the architecture, not an afterthought — not AI bolted onto an existing product. The primary design question is how to shape that relationship so both operate at their strongest.
Amplify, don't automate
The analyst stays in the loop; the agentic layer plugs in at specific points as a force multiplier. The goal is a sharper analyst, not an absent one.
Three systems, used where each wins
Humans, deterministic code, and the agentic layer each carry real strengths and real weaknesses. Most of the work stays deterministic code — the agent is used only where it earns its place, and its weaknesses are contained everywhere else.
Only ever make things as complex as they need to be.
The rule that governs every decision across all nine systems — how much context an agent gets, how much machinery a job deserves. Match the need, then stop.
There is no universal harness — which pieces it needs depends on what the agent is for. For a defensive-security agent, the terrain is nine core systems:
Data–Agent Interface
how telemetry is shaped for agents, and consumed
Shared State
findings persist across agents and hunts
Tools
external processes, mediated by the harness
Orchestration
arranging multiple agents, with gates
Skills
tacit hunting knowledge, made executable
Context Engineering
inference-time knowledge, provisioned
Feedback Loops
the system reflects, and improves itself
Evaluations
change measured, not vibes
Adversarial Resilience
the system defended from attacker influence
If you want to learn the field in practice — building the system end to end — the course is where to do it.