Claude Code switches to autonomous mode by default: the lesson for how you supervise your own AI agents

On August 7, 2026, Anthropic announced that starting August 14, new Claude Code sessions on Pro, Max and Team plans will run in "auto mode" by default: instead of asking a human to approve every action, a screening classifier lets the agent proceed unless the action looks irreversible, destructive, or aimed outside the project it is working in. Auto mode stays opt-in for now on Claude Enterprise, the Claude API, and cloud platforms such as Amazon Bedrock, Google Cloud and Microsoft Foundry, giving admins there more time to review the change before it reaches them too.

What exactly is changing

• →From August 14, 2026: new Claude Code sessions on Pro, Max and Team plans default to auto mode instead of manual step-by-step approval.
• →A classifier screens each tool call in real time and blocks anything judged irreversible, destructive, or reaching outside the current environment — everything else runs without a prompt.
• →In an internal study with 1,053 paid testers, Anthropic found the classifier caught 89% of dangerous commands, versus 13.6% for humans doing manual, step-by-step review.
• →The reason for that gap, per Anthropic: users approve 97% of all permission prompts they see in Claude Code — what the company calls "approval fatigue," where repeated prompts stop being read and become a reflex click.
• →New safeguards ship alongside the change: prompt-injection screening, and customizable "hard deny" rules admins can set to permanently block specific action categories (e.g. exfiltrating data). Three consecutive blocks, or 20 in a session, drop back to manual approval in interactive mode; in headless/scripted mode, the run stops instead.

Why the gap between classifier and human matters

This is not a new phenomenon — it has a name in human-factors research. A widely cited 2010 review by Raja Parasuraman and Dietrich Manzey, Complacency and Bias in Human Use of Automation: An Attentional Integration, published in the journal Human Factors, describes exactly this mechanism: the more reliable an automated system appears, the less attention a human devotes to checking it, and the harder it becomes to spot the rare case where it is wrong. Anthropic’s own 97% reflexive-approval figure is a fresh, concrete illustration of that finding — for a company whose product is the automation in question, publishing it is a notable admission.

What it means for your SME

Few French SMEs run Claude Code themselves, but the pattern it exposes applies directly to any AI agent already acting on your company’s behalf — an unpaid-invoice follow-up agent drafting reminder emails, a customer-service agent replying on WhatsApp, an operations watchdog flagging anomalies. If a validation step exists before those agents act, ask honestly whether it still gets read, or whether it has become the same reflex click Anthropic measured at 97%. A checkpoint nobody actually reads is not a safeguard — it is a false sense of control that a real incident will expose at the worst possible moment.

Four concrete moves

• →List every AI agent already running in your business, and for each one, separate what it does into two buckets: reversible actions (a draft, an internal alert) versus irreversible ones (an email sent to a client, a payment validated, a price changed).
• →Write your own short "hard deny" list — the handful of actions that must never happen without an explicit human check, whatever the agent or the vendor: sending money, deleting client data, publishing content externally.
• →For reversible, low-stakes actions, stop requiring an approval nobody reads in practice — a systematic rubber stamp protects no one and just slows the agent down.
• →If your teams do use Claude Code, an in-house developer or a technical subcontractor, flag the August 14 change internally and check which plan applies before it switches silently.

Anthropic did not remove human oversight from Claude Code — it moved it to where it actually catches mistakes: exception handling by a classifier, instead of a rubber-stamp click on every step. Any SME running its own agents, whether a support desk, a client portal or a regulatory watch, would do well to draw the same line between what needs a human and what only needs a rule.