Claude

How Claude Code Works

Claude Code runs on an "agentic" loop that gathers context, takes action, and verifies results. In this video, we break down the loop, the context window, tools, and permission modes so that Claude Code doesn't feel like a "magic box". Take the full course: claude.com/courses

May 14, 2026·2m·Watch on YouTube ↗

📖 CHAPTERS

1. 1

   Claude Code as an agent, not a chat window

   The video frames Claude Code as fundamentally different from typical chat apps because it can take actions in your environment, not just produce text. The core mental model is an “agentic loop” that repeatedly plans, acts, and verifies until the task is done.

2. 2

   Step 1: Prompt → context gathering

   After you enter a prompt, Claude Code first gathers the context it needs to complete the task. This sets up the rest of the loop by pulling in relevant conversation, files, and other signals.

3. 3

   Step 2: Model interaction → text or tool call

   Claude Code interacts with the model, which can respond either with plain text or with a tool call. Tool calls are the mechanism that lets the agent move beyond suggestion into execution.

4. 4

   Step 3: Taking action in the environment

   When a tool call is produced, Claude Code can take concrete actions like editing files or running shell commands. This is where Claude Code behaves like an agent operating within your workspace.

5. 5

   Step 4: Verification and loop repetition

   Claude Code checks whether the results match the original goal and are verifiable. If not, it re-enters the loop—gathering more context, acting again, and re-checking—until completion.

6. 6

   Human-in-the-loop steering during execution

   Even while the loop is running, you can influence outcomes by adding context, interrupting, or steering the agent. This keeps the workflow collaborative rather than fully autonomous.

7. 7

   Managed context window and automatic compaction

   Claude Code relies on a context window that limits how much conversation, file content, and command output it can retain. When the limit is reached, it compacts the conversation by removing or summarizing information to stay within bounds.

8. 8

   Tools as the backbone of agent behavior

   Tools are presented as the key enabler for agents: they let the system decide when to execute operations to make progress. Examples include reading files or searching the web, moving beyond purely conversational assistants.

9. 9

   Semantic search for deciding tool usage

   Claude Code uses semantic searching to determine when to call a tool and retrieve its output. This helps it pick relevant tools and information based on meaning rather than exact keywords.

10. 10

    Permission modes and safety tradeoffs

    Claude Code includes configurable permission modes that control when it must ask before editing files or running commands. The video emphasizes caution: loosening permissions can make mistakes harder to catch before they occur.

11. 11

    Putting it all together: agentic loop + context + tools + permissions

    The closing ties the concepts together: Claude Code combines an agentic loop, managed context window, tool execution, and configurable permissions in the terminal. This enables it to read a codebase, take action, and verify work—making it fundamentally different from a chat interface.