Skip to content
← The agents learning route

LOCKEDIN / VISUAL SYSTEMS LESSON

What actually
makes an agent?

Imagine a colleague who can ask for the next useful piece of information. An agent puts a model in that decision loop, while your application controls what it can do. Follow one task through three designs and see the difference for yourself.

Free lesson · No account or API key · Original, fictional teaching cases

01 / UNDERSTAND THE DIFFERENCEONE TASK · THREE DESIGNS

Who chooses the next step?

That is the difference to watch. Select a design, then follow the work one step at a time.

WHAT IT IS

Single model response

A language model generates a response from the information supplied to it. In this example, it has no connection to the ticket system or the approved guide.

A FAMILIAR COMPARISON

Like asking a colleague to help phrase a reply without giving them access to the case file. They can help with language, but cannot verify the missing facts.

“Chatbot” describes an interface as well as a product category. Here we compare a single model response with a workflow and an agent; all three could appear inside a chat window.

THE SAME REQUEST / SCRIPTED TEACHING CASE

Why is ticket T-104 blocked? Draft a reply with the approved next step. Do not send it or change access.

Step 1 of 3 · Guide available. The application passes along the request
INSIDE THE SYSTEMStep 1
Context →← Text response

→ Request → model response → person

Human reviewDraft or explicit request for help
END HERE

Select a component to see its responsibility. The highlighted component owns the current step.

Your application

The application passes along the request

The application sends the person's text to the model. The ticket ID is just text here; the model has not read a ticket because the text mentions its ID.

EXAMPLE MESSAGE / DATA
Why is ticket T-104 blocked? Draft a reply with the approved next step. Do not send it or change access.
WHO DECIDES?

Application code chooses to make one model call.

Authored replay · Fictional ticket and policy · No live model, SDK, tool call or external action. These frames show intended behavior that you must implement and test in a real build.

When to use this design

Useful when the needed information is already supplied. A convincing answer is not proof that the model has looked anything up.

Inspect the source records and full replay

Fictional source records

T-104 · workspace_access
Status: awaiting review
Reason: no approver assigned
GP-7 v3 · Workspace access
Ask the team lead to nominate an approver.
Keep access unchanged until the review is complete.

Single model response · Guide available

  1. The application passes along the request — Your application

    The application sends the person's text to the model. The ticket ID is just text here; the model has not read a ticket because the text mentions its ID.

    Why is ticket T-104 blocked? Draft a reply with the approved next step. Do not send it or change access.

    Application code chooses to make one model call.

  2. The model has words, not the case file — Language model

    The intended response asks for the missing information rather than pretending to know why this ticket is blocked. This is a scripted example of desired behavior, not a guarantee that every model will behave this way.

    I cannot see T-104 or your approved guidance. Share the relevant record and procedure, and I can help draft a reply.

    The model chooses the wording. It cannot run a lookup.

  3. The person has to supply the evidence — Human reviewer

    Changing the guide's availability changes nothing in this path: there is no tool connection. A chat interface can also sit on top of a workflow or an agent; the interface itself does not tell you the architecture.

    Next: obtain the ticket and approved guidance, then review any draft against them.

    The person must obtain the missing information.

02 / PLACE THE TECHNOLOGYDIFFERENT JOBS IN ONE SYSTEM

Where do the names fit?

A framework is reusable software you build with. A runtime carries out the work. A protocol defines how programs exchange messages. These names describe different responsibilities.

ASSEMBLE THE AGENT

LangChain

Software framework

Model integrations, tool definitions and a ready-made agent loop. It helps assemble the model-and-tool part of the system.

Build your first LangChain agent →
COORDINATE THE WORK

LangGraph

Orchestration framework + runtime

Connect steps, carry state, branch and pause for review. A graph can contain fixed steps, agent loops or both; it is not a new programming language.

Build your first LangGraph workflow →
CONNECT A CAPABILITY

MCP

Communication protocol

Connect an application’s MCP client to a server offering tools and data. It can be one connection beneath a workflow or agent, rather than the thing choosing its next action.

Inspect the connection below ↓
Look inside the tool connectionMCP in context

A tool can be a local function or use an ordinary API. If you connect it through Model Context Protocol (MCP), the application’s client exchanges structured messages with a server. The model does not speak directly to the database.

  1. 01 / APPLICATIONHost + MCP client

    Discovers the tool and sends a permitted request.

  2. 02 / CONNECTIONMCP server

    Validates the call and enforces access to the capability.

  3. 03 / SOURCETicket service

    Reads the authorized record. The result travels back.

MCP standardizes this exchange. It does not choose the workflow, grant access by itself or replace the model. This connection is an illustration; the replay above makes no MCP calls.

Build a real MCP client and server in the free lab →
03 / TRANSFER THE IDEAYOUR TURN

Choose the architecture for the work.

A good FDE can explain why the simpler design is enough—or what makes the more flexible one worth testing. Try each requirement before reading the reasoning.

Every ticket needs its record, the guide for its request type, and a draft for review. The order and branches are known.

Which would you investigate first?

Your selections and notes stay in this tab until you leave. Switching requirements keeps each draft here. Download anything you want to keep. No learner progress is recorded.

04 / TAKE IT INTO CODE

Build the loop you just explained.

The free workbook includes a real LangChain agent loop with a scripted model, a LangGraph workflow, and an MCP client/server lab. Run them locally, inspect their behavior, then change a condition. Paid provider calls are not required for these exercises.

Open the first LangChain build →

Already comfortable with the basics? Continue to the agent engineering course → Academy uses your existing training access.

Primary references behind this lesson

Original LockedIn Labs explanation and fictional case; vendor documentation is linked for further study. References checked 22 October 2026.