Claude Code Reins In Rogue Agent Drift

Lachlan Reed

So I- I was sitting at the bench yesterday, trying to push this tiny, tiny little CSS tweak, right? Just a, a margin fix, and suddenly my terminal goes absolutely wild. It's like, uh... it's like Claude decided to spin up a whole integration suite and- and run a security audit without me even asking. I was literally just sitting there watching my token balance burn.

James Turner

Oh man, the classic, the classic rogue agent spiral. But, you know, that is actually exactly what Anthropic is putting a stop to with the new Claude Code update that dropped on July nineteenth.

Lachlan Reed

Wait, really? July nineteenth? What- what did they actually change?

James Turner

Yeah, so in version 2.1.215, they completely pulled back on that unprompted behavior. The- the big headline here is that Claude no longer runs the /verify and /code-review skills on its own.

Lachlan Reed

Oh, thank goodness for that. Seriously, it was like working with a super eager junior engineer, you know? Like, you ask them to fix one typo on a button, and they, uh... they decide to independently trigger a multi-stage integration test and a container vulnerability scan. You're just sitting there, tapping your teacup, waiting for the cloud bill to arrive.

James Turner

Right! Exactly. You get a massive context-window bloat and this huge execution lag for a- a one-line change. It's too much. But now, it's- it's entirely manual. The developer is back in the driver's seat. If you want to test, you have to explicitly type /verify to run your local test scripts. It won't just guess anymore.

Lachlan Reed

Which, I mean, it makes a lot of sense. But... okay, if we have to do it ourselves now, how does the, the workflow actually change? Like, what's the playbook?

James Turner

Well, the power move now is using the flags. So, if you're in an interactive session, you can run /code-review --fix. That'll look at your local git diff, find the issues, and actually apply the fixes straight to your working tree.

Lachlan Reed

Oh, nice. So it still does the heavy lifting, you just have to give it the nod first.

James Turner

Precisely. And if you want to go, like, full enterprise mode, there's /code-review ultra. That actually dispatches a whole fleet of remote reviewer agents on Anthropic's cloud to do some deep, verified bug-hunting. It's pretty wild.

Lachlan Reed

Remote fleet, eh? That sounds like a- a proper way to burn through some serious compute, but at least now we're the ones pulling the trigger. But, you know, this does mean we've got to be a bit more disciplined. If we don't manually run /verify before we push... well, that safety net is gone. We could easily deploy broken code if we get lazy.

James Turner

Oh, totally. The training wheels are off. But hey, they did add some great quality-of-life stuff in the background too. Like, long-running tool calls finally have a progress heartbeat now, so the terminal doesn't just look frozen while it's thinking.

Lachlan Reed

Ah, yes! The "is it dead or is it just contemplating life" problem. That's a massive win. And hey, didn't they fix that Linux bug too? The one where if you ran a broad pkill -f inside the bash tool, it would... uh, what did it do? Target its own CLI?

James Turner

Yeah, it would literally commit suicide. It would kill Claude's own Node CLI runner process. That is finally fixed in the v2.1.214 patches. No more accidental self-destruction.

Lachlan Reed

Brilliant. Well, it's good to have the reins back. I'm off to manually verify some CSS. Catch you later, mate.

James Turner

Sounds good, talk soon.