The essay’s surface complaint is social — stop pasting Claude’s raw output into Slack and pull requests and calling it a contribution — but the engineering lesson underneath is about where judgment actually lives in an agentic workflow.
When you relay an LLM’s answer verbatim, you’ve inserted yourself as a proxy that adds latency and subtracts accountability. The recipient could have prompted the model themselves, faster, with context they control. What you were supposed to add — reading it, validating it, catching the plausible-nonsense — is exactly the step you skipped. The author’s example lands hard: a line like “NATS control-plane events: stream leader election / R3 quorum re-form during pod churn” reads as authoritative and might be entirely wrong, and the person forwarding it can’t tell which.
This maps straight onto how I think about human-in-the-loop design. The point of putting a person in an agent loop is that they’re a verification node, not a relay. A reviewer who pastes ticket text into a coding agent, ships the diff unread, then feeds reviewer comments back into the same agent hasn’t reviewed anything — they’ve added a hop. The implementation got done by two agents talking through a human who understood neither side.
The uncomfortable part: this is quietly what a lot of “AI-assisted” work already is. The value a human adds isn’t proximity to the model — everyone has that now. It’s the willingness to read the output, understand it, and re-express it in a form that certifies the reading happened. That certificate is the whole job.
The post is short and worth reading before your next PR, and the HN discussion is full of people recognizing themselves on both sides of it.
If your only contribution to a thread is forwarding a model’s output, what stops the recipient from cutting you out of the loop entirely?