Harness

A language model, on its own, only produces text. What turns Claude — the model — into an agent that reads files, runs commands, and opens a pull request is the harness: the program that wraps the model, hands it a set of tools, and executes whatever it decides to do. Claude Code is that harness. Anthropic's documentation uses the term literally — agentic harness — for the runtime that stitches model, tools, and execution environment into a capable agent.

#The agent loop

The harness runs a loop. On each turn it gathers context (reads the files it needs), takes action (edits, runs a command), and verifies (runs the tests, checks the output) — then decides the next turn based on what it just learned. The model reasons and picks the step; the harness executes it and feeds the result back for the model to read. It repeats until the task is done. The phases blend: the agent doesn't follow a fixed script, it adapts the sequence to each discovery.

#The tools

The model never touches your system directly. It asks, and the harness executes through a set of tools: Read and Write for files, Edit for targeted changes, Bash for the terminal, Glob and Grep for search, WebFetch and WebSearch for the internet, Agent to delegate to a subagent. Every extension you saw in this guide — skills, MCP servers, hooks — enters through this same mechanism: it widens the repertoire of tools the harness can run.

#The permission modes

Between "the model asked" and "the harness executed" sits a permission layer. It decides what runs on its own and what waits for your OK, and the mode tunes how strict it is:

• default — asks for confirmation before every action that changes something.
• acceptEdits — accepts file edits automatically, without asking for each one.
• plan — only reads and plans; changes nothing until you approve the plan.
• bypassPermissions — skips the confirmations: fast, but you give up the review. Reserve it for isolated environments.

Shift+Tab cycles the modes mid-session, so you tighten or loosen the reins depending on how much you trust the task.

#Closing the map

Understanding the harness stitches the rest of the guide together. CLAUDE.md provides the instructions, the extensions provide the capabilities, the hooks provide the guarantees — and the harness is what wraps all of it around the model, turn after turn of the loop, with the permission layer deciding what gets through. It's the difference between a chatbot that answers and an agent that does.

CLAUDE.md — the instructions the harness injects into the model on every turn of the loop Hooks — how to bind deterministic code to the events of the harness loop