Something shifted this month and I want to get it down before I forget I noticed it.

Anthropic shipped Claude Tag on the 23rd — Slack-native Claude, proactive, multiplayer, “Claude Code under the hood,” and 65% of Anthropic’s own product team’s code already comes from it. Same week, GLM 5.2 showed up handling ~80% of coding tasks at a fifth of Opus’s cost. The week before, dynamic workflows and nested subagents went GA in Claude Code itself. Cursor, Aider, Cline, Codex, Copilot — every plugin in the marketplace is racing in the same direction.

The shape of it: generating code is becoming a commodity. Fast. The price keeps dropping. The quality keeps converging. And the major AI companies are absorbing what used to be plugin-shaped opportunities — Slack agents, orchestration, monitoring, artifacts. If you’re building a plugin to make AI generate code, you’re now competing with Anthropic’s own roadmap.

I’ll be honest — I haven’t even used Claude Tag yet. I might be reading the signal too strongly. But the broader trend doesn’t depend on any one launch. Generation keeps getting cheaper. Trust doesn’t.

What I noticed about my own projectLink to heading

I have a Claude Code plugin called mx-workflow. Twenty-nine commands, twelve agents, a docs site. I’d been calling it “a Claude Code plugin for the full dev lifecycle.”

When I looked at the agents honestly, only four of the twelve actually generate things. The other eight interrogate generated code — silent failures, hallucinated APIs, suppressed errors, weak type design, doc rot, unjustified shortcuts. None of them write code. They look at code somebody (usually Claude) wrote and push back.

That’s not the headline I’d been selling. But it’s the half that actually does the work.

And it’s the half that doesn’t get commoditized as generation gets cheaper. If anything it gets more valuable — because when you’re shipping more AI-generated code, the cost of an undetected silent failure goes up, not down.

What I changed this weekLink to heading

I gave up trying to be the “ship faster” plugin and pointed mx-workflow at what it was already doing well.

The README now leads with “the review-grade quality layer for AI-generated code.” I shipped /mx:review as the starter dose — runs all eight review agents on changed code, hands back a verdict. Then five more refusal-grade commands behind it: /mx:hallucination-check for invented APIs and fabricated function signatures, /mx:second-look for reviewing other agents’ output (Claude Tag’s PRs, Cursor’s diffs, Copilot suggestions), /mx:reject as a CI gate, /mx:ratchet so quality only ever goes up, /mx:audit for comprehensive cron-friendly runs.

Four of the agents got pulled out as standalone Claude Skills too, so someone can install just one piece before committing to the whole plugin.

The build commands stay. They just stop being the thing on the front page.

HopefullyLink to heading

I might be wrong about all of this. I might come back to this post in six months and find that Claude Tag never landed, or that GLM 5.2 didn’t matter, or that some other shape took over the market that I’m not seeing yet. That’s fine — the changes are reversible. And I wrote down the trend I’m tracking — claim by claim, with quotes and dates — so I can audit my own thinking later as the signal moves.

But sitting where I am right now, reading what I’m reading, this feels like the right side of the bet. Generation keeps getting cheaper. Trust doesn’t. I’d rather be on the side of the business that makes AI-generated code trustworthy than the side that makes more of it.

Hopefully the new mx-workflow lands. Hopefully other people find it useful. Hopefully I’m reading the trend right.

If I’m not, I’ll write that post too.