Don't become a wrapper

Remember when we made fun of startups that just wrapped a prompt and an API call to OpenAI in a good-looking UI? We agreed those "ChatGPT wrappers" were cooked.

As hard as it is to say this, I think the same critique now applies to people.

I'm noticing a trend of engineers who've made themselves into "Agent wrappers". They prompt Claude Code for the plan, have it write the code, push up 50 PRs a day, paste whatever the agent generates into Slack, and repeat — until, if you took the agent away tomorrow, you would have to wonder what remains.

Once you start looking, you see it everywhere. Someone opens a 1,600-line PR and can't justify why a file exists. Slack messages start sounding like LinkedIn posts. Architecture decisions happen because the agent recommended them, and when you push on the reasoning the answer is some shape of "the model said so and it's smarter than me."

I get the appeal — the output is good enough that the failure modes don't bite right away. Shipping 10x+ the volume you did a year ago feels great. Every company all-hands has someone telling you to "use AI more," and few forces pull you back toward using your own brain too.

You learn where the cliffs are by walking the terrain yourself. An agent can accelerate that learning, but it cannot replace the need to understand and own the result. Without that ownership, wrapper-humans quickly become slop-cannons, defending the work with reasoning the agent invented for them and giving little thought to what happens to the team and systems around them.

Meanwhile, the alpha of being human is shifting under everyone. Some of the skills that made software engineers valuable in 2022 are commodities now. Agents can already produce syntactically correct code faster than most people. What remains scarce is taste, judgment, and context — reading a diff and instantly spotting that it's solving the wrong problem, missing a constraint, or working around a bug instead of fixing it.

I worry that, in the name of AI adoption and token-maxxing, wrapper-humans are slowly becoming the hands and eyes that fill today's missing connectors and harnesses — work the agents themselves may soon absorb.

The standard isn't whether you used an agent. It's whether you can explain its choices, reject them when they're wrong, and take responsibility for what ships.

So, think. If anyone with the same agent gets the same output as you, what remains defensible about you?