A colleague argues that a browser chat assistant could refactor your repository if it were given the same model Claude Code runs, and that the only real difference between the two is model quality. Where does the argument break down?

The principle — model quality and the ability to act are independent. Improving one does not supply the other.

Why the key is correct — Anthropic states the dependency plainly: tools are what make Claude Code agentic, and without tools Claude can only respond with text. A tool is an action it can take — read a file, edit code, run a shell command, search the web, spawn a subagent — and each use returns a result that informs the next decision. That last part is what makes a refactor possible at all, because a refactor is not one action but a chain of them, each shaped by what the previous one revealed. The surrounding layer has a name: the harness is the tools, context management and execution environment that turn a language model into a capable coding agent, with Claude Code as the harness and Claude as the model inside it. Give a browser chat window the same weights and you have changed the model and none of the harness, so the ceiling on what it can do is unchanged.

Why the others are wrong — attributing the gap to a coding-specific model predicts that swapping weights into a chat window would grant it file access, which it does not; the model is the same in both places. Blaming diff rendering mistakes a presentation detail for the blocker, when nothing there can open a file to diff. And enlarging the context window solves capacity, not capability: a window that could hold every file still offers no way to read one, edit it, or run the suite afterwards.

Remember this — ask what the assistant is connected to, not how clever it is. Connection is what decides whether anything can happen.

Sources — Anthropic's How Claude Code works and the Claude Code glossary.