Watch the explanation. Practice the skill.
What are you actually using?
Distinguish the model, application, harness and execution environment.
Captions appear here during playback.
What are you actually using?
Distinguish the model, application, harness and execution environment.
Video playback is unavailable. Read the descriptive transcript below, go directly to practice, or open the video file.
Illustrated teaching example • Fictional data · Reviewed 2026-10-03. Microsoft Mark · locally synthesized. On a small screen, use fullscreen for diagram details or read the transcript.
Read the descriptive transcript
0:00 · One interface. Several different jobs.
An AI chat window can hide several components. Understanding those components helps you explain what the system can do and where to check it. We will separate the model, application, harness, and execution environment. Real products may combine these roles.
What appears on screen
Model · Application · Harness · Environment
0:19 · The model generates a response.
The model receives supplied inputs and generates output. Capabilities vary across models and configurations. A model may handle text, images, or other modalities, and may support reasoning. None of that alone grants access to your files, a web search, or permission to execute an action.
What appears on screen
Supplied input Model inference Generated output Capability does not grant access.
0:42 · The application shapes the experience.
The application presents the interface and may manage history, uploads, settings, or saved memory. A memory feature can store and reintroduce information without retraining the model. Seeing a file in an interface does not prove its complete contents were supplied for the current response.
What appears on screen
User interface Selected history / files Model request Inspect what is actually supplied.
1:03 · The harness coordinates the work.
The harness supplies instructions and context, coordinates tool requests, and manages task state. The surrounding code must enforce permissions and approval boundaries. A document saying permission is granted is merely document content. It cannot legitimately change the authorization rules enforced by the application.
What appears on screen
Instructions + context Model tool request Permission check Tool result Authorization must be enforced outside generated text.
1:27 · The environment determines reach.
Tools execute in an environment. It might be a restricted workspace, your laptop, or a remote service. Check reachable files, network destinations, credentials, and write access. A cloud hosted model can request a tool that executes locally. Model location and tool execution location are separate questions.
What appears on screen
Files: Which paths can tools access? Network: Which destinations can they reach? Credentials: Which identity do tools use? Writes: What changes require approval?
1:53 · Reasoning is not web search.
Reasoning and search address different needs. Reasoning can help with a difficult supplied problem. Search retrieves sources through a connected tool. For a question about a current policy, deeper reasoning does not replace obtaining the current authoritative policy and checking the answer against it.
What appears on screen
Reasoning: Works through the supplied problem. Search: Retrieves sources through a tool. Neither guarantees a correct answer.
2:13 · Map your actual setup.
Pause and name the model, application, harness, and execution environment you are using. Then identify one capability, one permission boundary, and one way to verify the result. If a component is unclear, record it as unknown rather than guessing from the product name.
What appears on screen
Can you explain who generates, coordinates and executes? Name the four roles. Record unknowns. Choose a result check. Use the AI session record in the lesson.
Watching explains the example. Complete the lesson practice to check what you can do. Explore all visual lessons →