Agentic#
Agentic engineering is the discipline of composing LLMs into autonomous loops you can put to work. The model is a component, not the system. Around it build a prompt contract, a context window curated through retrieval, a tool-use layer, a memory layer, and an evaluation harness that keeps the loop honest as the underlying model changes. The mission outputs this discipline produces are the same outputs the operator would otherwise hand-roll, enrichment, summarisation, classification, code review, exploit drafting, defender simulation, but at the pace of a continuous loop rather than a per-task script.
This section is the builder’s angle. The chapters below cover the engineering primitives, the model SDKs, prompt patterns, retrieval, agents, local runtimes, and evals. Where agentic systems show up elsewhere in the handbook.
LLM-Assisted Analysis, LLM-as-analyst patterns for mission analysis (summarisation, entity extraction, RAG over collected corpora).
NLP, the broader NLP stack the LLM layer sits on.
Data Science, classical ML, anomaly detection, classification.
AI Systems as Targets, AI systems as a target class. prompt injection, model extraction, training-data poisoning.
Agents, the agents that run on the operator’s box (Claude Code, Aider, OpenHands, terminal agents).
Vendor SDKs for frontier models. Messages, tools, streaming, structured outputs.
Inputs that condition the model. Version-controlled, tested, reviewed like code.
Retrieval-Augmented Generation. Pull relevant documents,
concatenate, generate. The pragmatic answer to my data.
The loop. Model proposes a tool call, the app executes, the result feeds back.
Models running on hardware the operator controls. For sensitive corpora and air-gapped labs.
Tests for a model-driven system. Without evals, prompt changes are vibes-driven.