Systems Note // 04  ·  July 2026  ·  SGI Research

What Is An
Agentic AI?

Four systems that all look autonomous from the outside: LLM chatbots, RPA, RAG, and the real agentic loop. Only one of them can actually plan, act, and revise itself. Call each one below before you find out which.

↓ Start the diagnostic ↓

LLM Chatbots◆ RPA◆ RAG◆ Memory◆ Tools◆ Planning◆ Feedback◆ LLM Chatbots◆ RPA◆ RAG◆ Memory◆ Tools◆ Planning◆ Feedback◆

The Diagnostic

Agentic Or Not?

Answer first, then the real architecture behind that system opens up right below your guess.

The one test apply this to every card below
Does the output loop back and change what it does next? Or does it just end?
Dead end
Loops back both ways
01
LLM Chatbot

In plain English: you type a question, it answers from what it already knows. One question, one answer, done.

Query → System Prompt → LLM → Output, dead end
Querywhat you type
Just raw text typed in — no memory of earlier turns unless the system pastes it back in for you.
System Prompthidden setup instructions
Set once by whoever built this, invisible to you. It's why the same model can act like a pirate or a lawyer depending on what it's told to be.
LLMthe AI model
Predicts the next word over and over until it has a full response. It does not plan ahead or check its own work.
Outputthe answer
Once this prints, the system is finished. Nothing checks whether it's actually right.
02
RPA
Robotic Process Automation

In plain English: a script that repeats the same fixed steps every time, sometimes calling an AI along the way — but the steps never change on their own.

Query → Script Trigger → Execute Tools → Output, dead end
Querywhat you type
Just raw text typed in — no memory of earlier turns unless the system pastes it back in for you.
Script Triggerstarts the fixed steps
Fires the exact same sequence every time, regardless of what the query actually says.
Only Execute Toolsruns pre-set actions
Pure if-this-then-that automation — no AI reasoning in this branch at all.
LLMs (Use Tools)AI picks from a fixed toolset
The AI can only reach for tools someone already wired in ahead of time. It can't invent a new one mid-task.
Outputthe result
Whatever the script produced. If a step failed partway through, the script has no way to notice and adjust.
03
RAG
Retrieval-Augmented Generation

In plain English: before answering, it looks up real documents so the answer is grounded in facts, not just memory. It still just answers once and stops it never decides to go look up something else on its own.

Query → Embedding → Retrieve → Augment → LLM → Output, dead end
Querywhat you type
Just raw text typed in — no memory of earlier turns unless the system pastes it back in for you.
Embeddingturns your question into searchable numbers
A model turns your words into a list of numbers that captures meaning, so 'car' and 'automobile' land close together.
Knowledge Baseyour saved docs
A curated set of documents someone uploaded on purpose — manuals, internal docs, wikis.
Vector DBsearchable memory
Stores those number-lists so a search can find the closest matches in milliseconds instead of reading every document.
Web Searchlive web results
A live query out to the internet, same as typing into a search bar.
Augmentadds that info to the prompt
The retrieved snippets get pasted directly into the prompt before the model ever sees your original question.
LLMthe AI model
Predicts the next word over and over until it has a full response. It does not plan ahead or check its own work.
Outputthe answer
Once this prints, the system is finished. Nothing checks whether it's actually right.
04 · Final
The Agentic Loop

In plain English: it decides what to do, does it, checks whether it actually worked, and tries something different if it didn't on its own, without you asking again.

Orchestrator + Memory / Tools / Planning / Feedback → Output → loops back into Query
Drag to rotate real WebGL
AGENT_01
Core Loop
Orchestrator + 4 Capabilities
Orchestrator LLMDirects the loop. Reads the query, decides which capability to invoke, and arbitrates between memory, tools, planning, and feedback
MemoryRetains prior turns and intermediate results so decisions build on what already happened
ToolsExecutes external functions code, APIs, search chosen by the orchestrator rather than hard-coded in advance
PlanningBreaks the query into an ordered sequence of sub-tasks before any tool is called
FeedbackEvaluates the outcome of each step and revises the plan the mechanism a static pipeline lacks entirely
PROTOCOL
Multi-Agent Protocol
Coordination Layer
DiscoverIdentifies which specialist agents exist and what each one is capable of doing
Share TasksRoutes sub-tasks from the orchestrator to whichever specialist agent is best suited to handle them
Update Task InfoKeeps task state synchronized across agents as work moves back and forth through the loop
SPECIALIST_01 Coding Agent

Writes, edits, and executes code sub-tasks handed off by the orchestrator.

SPECIALIST_02 Retrieval Agent

Pulls grounding data from knowledge bases, vector stores, or the web on request.

SPECIALIST_03 Citation Agent

Traces claims in the output back to their sources before the answer is returned.

↻Output feeds back into Query the loop a linear pipeline never closes
✓
Diagnostic Complete 0/4
A script that calls tools isn't an agent. A loop that revises its own plan is.

A SCRIPT THAT CALLS TOOLS
IS NOT AN AGENT.
A MODEL THAT RETRIEVES CONTEXT
IS NOT AN AGENT.
ONLY A LOOP THAT REVISES ITSELF IS.

More Systems.
More Notes.