Clear definitions of the terms we build with: agents, agentic development, RAG, automation, and the systems that tie them together.
AI agent
An AI agent is software that uses a language model to decide and take actions in a loop, retrieving context, calling tools, and checking its own work, rather than only generating text.
Where a chatbot answers a prompt, an agent does things: it reads a situation, chooses an action, calls a tool to carry it out, observes the result, and decides what to do next. That loop is what separates an agent from a model behind a chat box.
Good agents pair four moves, retrieval, reasoning, tool use, and verification, with clear limits on what they can do autonomously and where a human steps in.
Agentic development
Agentic development is building software where an AI agent decides the next step at runtime, retrieving, reasoning, acting, and verifying, instead of following a fixed script.
Traditional automation encodes every step in advance. Agentic development inverts that: you give the agent goals, context, and tools, and it determines the path at runtime.
It trades some predictability for adaptability, so the engineering work shifts to retrieval quality, tool design, verification, and the human gates that keep autonomy safe.
Retrieval-augmented generation (RAG)
RAG is a technique that gives a language model your own documents at answer time, so it responds from retrieved sources, with citations, instead of guessing.
A base model only knows its training data. RAG closes that gap by indexing your sources, retrieving the most relevant passages for each question, and constraining the model to answer from them.
Quality lives in the retrieval layer, chunking, hybrid search, re-ranking, and evaluation, not in the prompt. Done well, it turns a guessing model into a grounded, traceable one.
Multi-agent system
A multi-agent system splits a large task across specialized AI agents behind an orchestrator, each with a narrow job, clear inputs, and a defined hand-off.
A single agent handles a bounded task well, but as scope grows one agent juggling everything becomes hard to reason about. Multi-agent systems decompose the work so each agent stays legible.
An orchestrator routes work between agents and keeps the overall flow observable, the same way a well-structured team divides responsibility.
AI automation
AI automation uses a language model to handle the judgment-heavy steps inside a workflow, classifying, extracting, summarizing, deciding, that rigid rules can't.
Rule-based automation excels at deterministic work: clear trigger, clear action. AI automation fills the gaps that need judgment, so a workflow runs end to end instead of stalling on the one step a human used to handle.
The discipline is to use rules for the deterministic spine and AI only where judgment is genuinely required, with confidence thresholds that escalate uncertain cases.
AI operating system
An AI operating system is what you get when many automations and agents share a common context layer and a single control plane, reasoning from one source of truth.
A handful of disconnected automations become their own maintenance problem: no shared context, no common oversight. An AI operating system composes them: a shared context layer and a single control plane so every action is observable.
It's the move from scattered scripts to a coherent system that compounds in value as you add to it.
Large language model (LLM)
A large language model is an AI model trained on vast text to predict and generate language; it powers chat, summarization, extraction, and the reasoning inside AI agents.
LLMs are the engine behind modern AI features, but on their own they only know their training data and will answer confidently even when wrong.
Production systems wrap them with retrieval (RAG), tools, and verification so their output is grounded, actionable, and checkable.
Vector embedding
A vector embedding turns text into a list of numbers that captures meaning, so a system can retrieve passages by semantic similarity rather than exact keywords.
Embeddings let a RAG system find the passages most relevant to a question even when the wording differs, by comparing vectors in a high-dimensional space.
They are the backbone of semantic search and the retrieval step in most RAG pipelines.
Human-in-the-loop
Human-in-the-loop is a design where an AI system escalates uncertain or high-stakes decisions to a person, using confidence thresholds, approval gates, and audit trails.
Autonomy is a dial, not a switch. The engineering question is which decisions are safe to automate and which need a human gate.
Human-in-the-loop keeps autonomy from outrunning oversight: confident, low-stakes actions run automatically; the rest are routed to a person with a clear audit trail.
AI harness
An AI harness is the engineering scaffold around a model: prompts, tools, retrieval, evaluation, and guardrails, that turns a raw model into a reliable, shippable system.
A model is not a product. The harness is everything that makes it dependable: structured prompts, tool integrations, retrieval, automated evaluation, and the guardrails that catch failures.
It's where most of the real engineering of an AI feature actually lives.