The Spectrum of GenAI Applications

AI system does not mean agent: separate knowledge, control flow, and actor topology

◷ Estimated reading time: 5 min

A GenAI system is any software system that uses a generative model as one of its runtime components. Some systems generate one answer. Some retrieve private knowledge. Some follow a fixed business process. Some let a model decide the next action. Only the last group is meaningfully agentic.

Three Axes, Not One Maturity Ladder

It is tempting to draw a single progression—chatbot, RAG, workflow, agent, multi-agent—and assume each step is a more advanced replacement for the previous one. That is useful as a first sketch, but architecturally incomplete. These labels answer different questions:

Axis	Question	Examples
Knowledge access	Where does the system obtain facts?	Model memory, long context, RAG, live tools
Control allocation	Who decides the next step?	Developer-defined workflow or model-directed agent loop
Actor topology	How many independent reasoning actors participate?	Single model/agent or multiple specialized agents

A customer-support system can therefore be a RAG workflow, a RAG-enabled agent, or a multi-agent system that uses RAG. RAG describes knowledge access; agenticness describes runtime control; multi-agent describes organization.

Common System Shapes

Direct model call: the application assembles input, calls a model, and returns or processes the output.
Chatbot: a direct model interaction with conversational history and a user-facing turn-taking interface.
RAG application: the application retrieves relevant external information and includes it in the model context.
Deterministic AI workflow: developers predefine the sequence, branches, retries, and approvals; models perform selected steps.
Agentic workflow: the model chooses part of the path at runtime, such as which tool to call or whether another step is required.
Multi-agent system: work is divided among multiple agents with distinct roles, state, or ownership boundaries.

One Requirement, Several Architectures

Suppose an employee asks, "Can I carry unused vacation days into next year?"

Architecture	How it handles the request	When it fits
Direct LLM	Answers from model knowledge	Low-stakes generic explanation; no company-specific truth required
RAG	Retrieves the current employee handbook, then answers with citations	The policy exists in documents and the flow is predictable
Workflow	Retrieves policy, checks location and employment type, then formats a governed response	Business rules and required steps are known in advance
Agent	Decides which systems to query, asks clarifying questions, and determines whether escalation is needed	The path depends on runtime evidence and cannot be fully enumerated
Multi-agent	An HR policy agent coordinates with payroll and legal agents	Independent domains or organizational boundaries justify separate actors

The most complex option is not automatically the best. Every increase in autonomy adds latency, cost, permissions, state, and failure modes.

Start With the Simplest Sufficient Architecture

A useful decision habit is to ask what uncertainty actually exists:

Can one model call solve the task safely?
Does the answer require private, current, or verifiable knowledge?
Is the execution sequence known in advance?
Must the model decide among actions at runtime?
Do separate specialist actors provide a real boundary or capability advantage?

Architecture ruleDo not begin with "How do we add an agent?" Begin with "What decisions cannot be encoded reliably in a simpler system?"

Key Takeaways

Not every GenAI system is an agent.
RAG describes knowledge access; workflow versus agent describes control; single versus multi-agent describes actor topology.
The same product may combine several patterns.
Use the simplest architecture that meets the requirement and risk level.

A system always retrieves three approved documents, sends them to a model, and formats the answer through a fixed validator. What is the best description?

1.1 From Traditional Software to GenAI Systems 1.3 What Makes a System Agentic?

loading…

The Spectrum of GenAI Applications

AI system does not mean agent: separate knowledge, control flow, and actor topology

◷ Estimated reading time: 5 min

Three Axes, Not One Maturity Ladder

Axis	Question	Examples
Knowledge access	Where does the system obtain facts?	Model memory, long context, RAG, live tools
Control allocation	Who decides the next step?	Developer-defined workflow or model-directed agent loop
Actor topology	How many independent reasoning actors participate?	Single model/agent or multiple specialized agents

Common System Shapes

Direct model call: the application assembles input, calls a model, and returns or processes the output.
Chatbot: a direct model interaction with conversational history and a user-facing turn-taking interface.
RAG application: the application retrieves relevant external information and includes it in the model context.
Deterministic AI workflow: developers predefine the sequence, branches, retries, and approvals; models perform selected steps.
Agentic workflow: the model chooses part of the path at runtime, such as which tool to call or whether another step is required.
Multi-agent system: work is divided among multiple agents with distinct roles, state, or ownership boundaries.

One Requirement, Several Architectures

Suppose an employee asks, "Can I carry unused vacation days into next year?"

Architecture	How it handles the request	When it fits
Direct LLM	Answers from model knowledge	Low-stakes generic explanation; no company-specific truth required
RAG	Retrieves the current employee handbook, then answers with citations	The policy exists in documents and the flow is predictable
Workflow	Retrieves policy, checks location and employment type, then formats a governed response	Business rules and required steps are known in advance
Agent	Decides which systems to query, asks clarifying questions, and determines whether escalation is needed	The path depends on runtime evidence and cannot be fully enumerated
Multi-agent	An HR policy agent coordinates with payroll and legal agents	Independent domains or organizational boundaries justify separate actors

The most complex option is not automatically the best. Every increase in autonomy adds latency, cost, permissions, state, and failure modes.

Start With the Simplest Sufficient Architecture

A useful decision habit is to ask what uncertainty actually exists:

Can one model call solve the task safely?
Does the answer require private, current, or verifiable knowledge?
Is the execution sequence known in advance?
Must the model decide among actions at runtime?
Do separate specialist actors provide a real boundary or capability advantage?

Architecture ruleDo not begin with "How do we add an agent?" Begin with "What decisions cannot be encoded reliably in a simpler system?"

Key Takeaways

Not every GenAI system is an agent.
RAG describes knowledge access; workflow versus agent describes control; single versus multi-agent describes actor topology.
The same product may combine several patterns.
Use the simplest architecture that meets the requirement and risk level.

A system always retrieves three approved documents, sends them to a model, and formats the answer through a fixed validator. What is the best description?

1.1 From Traditional Software to GenAI Systems 1.3 What Makes a System Agentic?