The Claude Code v2.1.215 release, published on July 19, brings a small but significant change: the agent stops launching /verify and /code-review on its own initiative. From now on, you invoke them when you want them to run.

Why this matters to you. Until now Claude Code could decide on its own to run a verification pass or code review while working, even without you asking. Now those steps only happen on command. If you use the agent in your daily work, it means fewer unexpected operations in the middle of your workflow and more control over when to stop and review the results. The thing to watch: if you were relying on the agent’s initiative for checks, now you have to remember when to trigger them.

It’s the latest step in a series of rapid patches we’ve been tracking since early July. Yesterday the v2.1.214 closed a Windows permissions bypass; today Anthropic shifts the agent’s behavior toward the model where it does what you explicitly ask, dropping the autonomous initiative on checks.

In detail

What /verify and /code-review do.

In Claude Code, «skills» are commands that launch structured procedures. /verify runs a check on the code produced by the agent: it verifies that tests pass and that the result matches what you asked for. /code-review runs a revision pass on the work done, flagging issues and parts to redo. They’re self-check tools available to the agent while it works.

What changes.

Before this release, Claude Code could launch these commands when it thought the time was right. If it was finishing a complex task, it could decide on its own to run a verification pass before handing over the result. With v2.1.215 this autonomous initiative is removed: the agent executes /verify and /code-review only when you type them.

It’s a shift in approach that touches the central issue of agent UX: how much it decides on its own versus how much it follows your explicit instructions. In recent months the industry trend has been toward increasingly autonomous agents that make decisions and take actions without step-by-step supervision. This patch goes in the opposite direction on a specific area, to give more control to the user.

The context of the patch series.

The v2.1.215 arrives after ten days of rapid releases. We’ve tracked the series from v2.1.208 on July 14 onward: fixes for blocked dialogs in background sessions (v2.1.209), visible timer for long operations (v2.1.210), fork as independent background session (v2.1.212), Windows permissions fixes (v2.1.214). The pace indicates that Anthropic is consolidating Claude Code for daily use, not adding experimental features.

Today’s change fits this trajectory. An agent that launches verifications on its own might look good in a demo, but in production it means wasted time when unnecessary and unexpected results when you least expect them. Making it explicit is a maturation move.

Knowledge limits.

The release note is one line long. We don’t know if autonomous initiative was removed because it caused concrete problems (verifications launched at the wrong time, loops, slowdowns) or as a preventive design choice. We don’t know if other commands will follow the same path, or if there will be an option to reactivate autonomous behavior for those who prefer it.

What we do know is the direction: Claude Code is moving toward a model where you decide when to check, and the agent executes the commands you give it.