How Claude Code Works

The Agentic Loop

• User enters a prompt
• Claude gathers required context by interacting with the model
• Model returns text or tool calls that Claude executes
• Claude takes action (e. g. , editing files, running commands)
• Claude verifies results against the original prompt goal
• If successful, waits for next prompt; if not, repeats the loop
• User can add context, interrupt, or steer the model during the loop

Context Window

• Stores conversation history, file contents, and command outputs
• Has a size limit
• When full, Claude automatically compacts the conversation by summarizing and removing non-essential information

Tools

• The backbone of how agents work
• Allow Claude to execute code rather than just output text
• Examples: read file tool, search web tool
• Claude uses semantic searching to determine when to call tools

Permission Modes

• Default: Asks explicit permission before editing files or running shell commands
• Auto-accept edits (Shift+Tab): Automatically edits files without asking, but still asks for commands
• Plan mode: Uses read-only tools to compile an action plan before starting
• Caution: Skipping permissions can make mistakes harder to catch

Core Difference from Chat Applications

• Combines agentic loop, managed context window, tools, and configurable permissions
• Can read codebases, take action, and verify its own work
• Fundamentally different from standard chat windows

Takeaways

• The Agentic Loop: Claude iteratively gathers context, executes tool calls (file edits, commands), verifies results against your goal, and repeats until successful—you can interrupt or steer at any point.

• Tools Enable Action: Tools are what allow Claude to execute code and take real actions (read files, run commands, search) rather than just output text; Claude uses semantic searching to decide when to call them.

• Context Window Management: The conversation history and file contents are stored in a limited context window that automatically compacts by summarizing when full.

• Permission Modes Control Safety: Default mode asks permission before edits/commands; auto-accept skips file edit confirmations; plan mode uses read-only tools first—choose based on your risk tolerance.

• Fundamentally Different from Chat: Claude Code combines agentic loops, managed context, tools, and permissions to read codebases, take autonomous action, and verify its own work—not just a standard chat interface.