Module 1: The Agentic Loop

The loop begins when Claude Code receives a prompt. This can happen in several ways:

Regardless of how the input arrives, it enters the same pipeline. The input method only affects what happens around the edges: whether the UI renders interactively, whether permission prompts appear, and whether the output streams to stdout or a terminal widget.

Before the prompt reaches the model, Claude Code constructs a rich context window. This is not just your message — it is your message wrapped in layers of environmental information that give the model situational awareness.

Two context blocks are prepended to every API call:

The result is a prompt that looks nothing like what you typed. Your simple question arrives at the model surrounded by hundreds of tokens of context that orient it to your project, your preferences, and your current state.

The assembled context hits the Claude API. The model analyzes the full prompt — your message, the system context, and the conversation history — and decides what to do.

This is where tool selection happens. The model does not just generate text; it can emit structured tool calls as part of its response. Each tool call specifies a tool name and parameters, such as "read the file at path X" or "run this bash command."

The model may decide to:

The key insight: the model is making autonomous decisions about what actions to take. You did not tell it which tools to use. It figured that out from the task description and the available tool set.

Before any tool executes, Claude Code checks whether it is allowed to run. Every tool has a checkPermissions method that returns one of three verdicts:

The permission outcome depends on the current permission mode and any rules defined in your settings. Read operations (like reading a file or running a search) are typically allowed. Write operations (like editing a file or running a shell command) often require approval.

This is a critical safety layer. The model can reason about any action it wants, but it cannot execute anything without passing through this gate. We cover the permission system in depth in Module 4: Permissions & Trust.

Approved tools execute in your local environment. This is real execution — files are written, commands run, searches performed — all within your terminal process. There is no remote sandbox. The tool runs, produces output (stdout, stderr, file contents, search results), and that output is captured.

The captured output becomes a tool result message that is appended to the conversation history. This is the observation that feeds the next reasoning step.

The model now receives the tool result and makes another decision: is the task complete, or is more work needed?

If more work is needed, execution jumps back to Step 3. The model reasons again, potentially calling another tool, and the cycle repeats. This continues until the model determines the task is finished and emits a final text response.

This is what makes Claude Code an agent rather than a chatbot. A chatbot responds once. An agent acts, observes, reasons, and acts again — as many times as necessary.

The default mode. You launch claude in your terminal and interact through a live UI with streaming output, permission prompts, and rich formatting. The full six-step cycle runs with all safety checks and user interaction points.

For automation and scripting. Input arrives via stdin or the --print flag. Output goes to stdout as plain text. Permission prompts are suppressed — only pre-approved tools run. This mode is useful for CI pipelines, git hooks, and shell scripts.

Claude Code can spawn child instances of itself. The parent agent delegates a sub-task to a new agent that runs in its own context, executes its own agentic loop, and returns the result. This enables parallel execution and task decomposition. We cover this in detail in Module 9: Multi-Agent Patterns.