Claude Code Ask Mode Is Now the Default: What Changed on August 14, 2026

On August 14, 2026, Anthropic quietly shipped a default change that affects every developer running Claude Code on a Pro, Max, or Team subscription. Previously, new sessions started in Auto mode — Claude would take actions (write files, run commands, execute destructive operations) without stopping for approval on each step. The new default is Ask mode, labeled "Manual" in the CLI since v2.1.200, which requires approval before every file write, command execution, or destructive action.

If you were running Claude Code every day, you noticed this immediately. The session that used to proceed autonomously through a multi-step refactor now stops at each edit and waits. That is not slower in theory — it is operating as designed — but it is a materially different experience from what most developers had built their workflows around.

What Ask mode actually means in practice

In Ask mode, every action that changes state on your filesystem or executes a command surfaces a prompt before it proceeds. You see exactly what Claude intends to do: which file it wants to write, what content it will produce, what command it is about to run. You approve or reject each step. Nothing happens without explicit sign-off.

The experience is not unlike reviewing a surgeon's plan before each incision rather than watching them operate through a window. It is more work per unit of delegation, but it means you have reviewed and understood each change before it happened — which matters when the codebase is one you own and will be maintaining.

Ask mode was already available before August 14. Many developers using Claude Code in production contexts had switched to it deliberately, particularly for work in critical paths, security-adjacent code, or repositories where an unexpected file change could cascade. The August 14 change makes it the baseline for everyone rather than an opt-in setting.

Why Anthropic changed the default

Anthropic has not published a detailed rationale for the timing, but the direction is consistent with several months of stated safety positioning around agentic systems. The company has consistently framed autonomous agent behavior as something that should require deliberate opting-in rather than being the default posture for new users or new sessions.

The practical case for this default is strong. Developers new to Claude Code sometimes misjudged how much autonomy they were granting when they started a session in Auto mode. An agent running in Auto on an unfamiliar codebase can produce cascading file changes that are individually reasonable but collectively wrong — a multi-file refactor that touches more than intended, a test cleanup that removes coverage the developer did not know was covering something else, a dependency update that silently introduces a breaking change.

With Ask mode as the default, every new user starts in a position where they understand and have approved every change before it lands. Switching to Auto is one setting change and takes ten seconds — but it is now an active choice rather than a passive one.

There is also an enterprise and compliance angle. For developers working in regulated industries or on security-sensitive codebases, having Ask mode as the default aligns Claude Code with the kind of audit-trail and explicit-approval workflows that compliance frameworks often require. Auto mode delegated that approval implicitly; Ask mode makes it a documented event.

How to switch back to Auto mode

If you want the previous behavior — Claude proceeding autonomously without per-step approval — switching back is straightforward. From inside a Claude Code session, run /config and change the approval mode setting to Auto. You can also set this in your global configuration so it persists across sessions:

# In your Claude Code session
/config set approvalMode auto

# Or in your ~/.claude/config.json
{
  "approvalMode": "auto"
}

Team and enterprise accounts can set the default at the organization level through the Anthropic console, overriding the per-user default. This means if your team has established workflows built around Auto mode, an admin can restore the previous behavior for the whole team without requiring each developer to change their individual configuration.

One note on terminology: the CLI uses "Manual" and "Auto" as the mode labels since v2.1.200, even though Anthropic's documentation and product UI still reference "Ask" mode. They are the same thing. The renaming to "Manual" in the CLI was an earlier change; the August 14 update changed which mode is the default, not the labels.

Whether to switch back

The more useful question is not how to undo the change, but whether you should. The answer depends on your workflow, your codebase, and your risk tolerance — and most developers who have thought about it carefully end up with a nuanced position rather than a blanket preference.

Arguments for keeping Ask mode as your default:

• You catch more mistakes earlier. Approving each file write means you see the shape of Claude's approach at each step. Problems that would have been buried in a diff are visible as they happen. This is genuinely valuable for anything touching business logic, data models, or API contracts.
• The cognitive overhead is lower than it feels. Reviewing a proposed file write and clicking Approve takes two seconds when Claude's plan is correct. The overhead only accumulates when you are rejecting or redirecting — which is exactly when you want overhead, because those are the moments that matter.
• It prevents accumulation of silent errors. Auto mode is most dangerous when Claude makes an early mistake and compounds it through subsequent steps. Ask mode breaks that failure mode by forcing a review before each compounding action.

Arguments for switching back to Auto:

• Well-bounded tasks with clear scope. If you are asking Claude Code to "add docstrings to every function in this module," the risk of each individual file write is low and the approval overhead adds friction without adding safety.
• Throw-away environments. In a sandbox, a branch built for quick prototyping, or a generated codebase you plan to review wholesale rather than incrementally, per-step approval does not buy much.
• High-trust workflows. If you have been using Claude Code for months, you understand its tendencies, you have CLAUDE.md instructions that constrain its behavior, and you know which types of tasks it handles reliably — Auto mode may be the right default for those known-safe task classes.

A practical split-mode strategy

The cleanest approach for most developers who do serious work with Claude Code daily is not a binary choice between Ask and Auto — it is a split strategy based on task risk class.

Start sessions in Ask mode (the new default). For bounded, low-risk tasks where you understand the scope completely, switch to Auto for that session or that task block. When the task finishes or you move to something with higher stakes, return to Ask. The overhead of switching modes is minimal; the discipline of making it a conscious choice rather than an ambient setting is the real value.

That framing also makes the August 14 default change easier to accept: Anthropic did not remove Auto mode or make it harder to access. They changed which mode you have to opt out of versus opt into. If you have been thoughtful about when you delegate fully to an AI agent and when you want a review step, this default change probably aligns with how you were already thinking about it.

What this signals about Claude Code's trajectory

The Ask mode default change sits alongside a broader set of August 2026 improvements to Claude Code's supervision ergonomics: the improved agent view with colored state labels and classifier-written headlines, the enhanced /doctor command for setup diagnosis and repair, and the multi-agent peek panel that shows exactly why a blocked agent is waiting.

The pattern is consistent: Anthropic is making it easier to operate Claude Code as a supervised collaborator rather than an autonomous black box. The August 14 default change is the most structural of these moves — it is not a UI polish, it is a philosophy statement about where responsibility should live by default.

For developers who have been treating AI agents as force multipliers for their own judgment rather than as replacements for it, that philosophy statement is comfortable. The default is now aligned with how experienced Claude Code users were already operating.