Claude Code made 'ask first' the default — so should you

In a quiet July release, Anthropic changed a default in Claude Code that says more about where agentic AI is headed than most product launches do. The leading AI coding agent — the thing that edits files, runs commands, and opens pull requests on its own — now asks before it acts, by default. For anyone thinking about turning AI loose on their business systems, that's the signal: the company most bullish on autonomous agents just decided the safe default is a human saying "yes" first.

What actually happened

Per the Claude Code changelog, version 2.1.200 (released July 3, 2026) "changed the 'default' permission mode to 'Manual' across the CLI, VS Code, and JetBrains." In plain terms: out of the box, the agent now requires explicit approval before it does anything consequential, instead of proceeding on its own. The same release also stopped AskUserQuestion dialogs from auto-continuing — the agent waits for a real answer rather than picking one and moving on. You can still opt into faster, less-interrupted modes; they're just no longer what you get by default.

This is a deliberate rebalance. For two years the pitch for coding agents was speed — let it run, review later. The new default inverts the burden: automation is opt-in per action, not assumed. Anthropic isn't backing away from autonomy — it's saying autonomy should be a choice you make on purpose, not a setting you inherit.

Why it matters for your business

You're probably not running Claude Code across your accounting stack. But you are, increasingly, being sold AI agents that take actions — send the email, update the record, issue the refund, post the order. The question every operator should ask before deploying one is the same question this changelog just answered: what is this thing allowed to do without me?

The right answer is almost never "everything" and almost never "nothing." It's a line drawn on purpose. Reads, drafts, and reversible steps can run unattended — that's where automation earns its keep. Anything that moves money, touches a customer, or can't be easily undone gets a human gate. That's not a lack of trust in the AI; it's the same control you'd put on a new employee with a company card. The industry's most aggressive agent vendor just made that gate the default. If your AI setup lets an agent take irreversible actions with no approval step, you're now running looser than Claude Code ships out of the box.

We build automations with those gates designed in from the start — unattended where it's safe, approval-gated where it's not — so you get the speed without handing over the keys.

Sources: Claude Code releases (GitHub), Claude Code permission modes docs.