← AI Knowledge Home
01 · GOAL
02 · PLAN
03 · ACT
04 · EXECUTE
05 · OBSERVE
06 · CONTINUE OR FINISH
09 · AGENTS, TOOLS & MCP
How does a model become an agent that can act?
A model generates responses. An agent can pursue a goal, use tools, observe results, and continue through multiple steps.
MODEL · intelligence+TOOLS · capabilities+AGENT LOOP · coordination=AGENT SYSTEM
THE PROBLEM THIS MODULE SOLVES
USER REQUEST
“Check order 123 and send the customer an update.”
WHY A MODEL ALONE IS NOT ENOUGH
A model can understand the request and draft text. By itself it cannot:
- query the order system
- execute an external action
- track workflow state
- ask for approval
- continue a multi-step task
That extra system is the agent layer.
THE BIG PICTURE
Agent Architecture
USER GOALWhat needs to happen?
AGENT SYSTEMCoordinates the work
MODELProvides intelligence
PLANNINGChooses next action
STATETracks task status
AGENT LOOP
OBSERVE → DECIDE → ACT → OBSERVE RESULT ↺
OBSERVE → DECIDE → ACT → OBSERVE RESULT ↺
TOOLSSearch · Files · Email · DB
MCP / APIConnection layer
EXTERNAL SYSTEMSCRM · ERP · Browser · Database
HUMAN APPROVALReview when required
SKILLS / PLUGINSPackaged capabilities
MULTI-AGENTAgents coordinate work
HOW THE AGENT LOOP WORKS
From a goal to a completed task
Agent chooses the next action.
Agent requests a tool.
Application performs the action.
Tool result comes back.
Update state, repeat, or ask approval.
ONE REAL EXAMPLE
Order-status update
SCENARIO · “Check order 123 and send the customer an update.”
USERCheck order 123
AGENTNeeds order status
TOOL CALLget_order_status(“123”)
APPLICATIONExecutes Order API
RESULT“Shipped”
STATE UPDATEStatus is known
AGENTDrafts customer email
HUMAN APPROVALApprove before send
FINAL ACTIONEmail sent / prepared
UNDERSTAND EACH LAYER
Different roles, one agent system
THINK
Agent · Planning · Agent Loop
Decides what should happen next.
ACT
Tool Calling · Tools · Skills
Does something outside the model.
REMEMBER
State · Memory
Keeps useful task information.
CONNECT
MCP
Connects AI applications to tools and resources.
CONTROL
Human in the Loop
Adds review or approval.
COORDINATE
Multi-Agent Systems
Lets multiple agents work together.
WHY THE LOOP MATTERS
One response is not an agent loop
ONE-SHOT MODEL
Input
↓
Output
AGENT LOOP
Goal → Action → Result → New decision → Action ↺
Agents can adapt after seeing intermediate results.
TOOL CALLING
A request is not the real-world action
MODEL
“I need order status.”
↓“I need order status.”
TOOL CALL
get_order_status({order_id: “123”})
↓get_order_status({order_id: “123”})
APPLICATION / RUNTIME
executes
↓executes
ORDER SYSTEM
↓RESULT
{status: “shipped”}
↓{status: “shipped”}
MODEL CONTINUES
KEY DISTINCTIONS
What not to confuse
MODEL vs AGENT
MODELGenerates or reasons.
AGENTUses model + tools + state + loop.
A model can answer. An agent can carry out a task.
TOOL CALLING vs TOOL EXECUTION
CALLINGRequests an action.
EXECUTIONActually performs the action.
Model: get_order_status(“123”). Application: calls the Order API.
STATE vs MEMORY vs CONTEXT
STATECurrent task status.
MEMORYInformation that may persist.
CONTEXT is information visible to the model right now.
MCP vs TOOL
MCPConnection standard.
TOOLActual capability.
MCP can connect an AI app to a file-search tool.
REMEMBER THIS
An agent is a controlled system around a model: it plans, acts through tools, observes results, keeps state, and loops until the goal is complete or a human must decide.