Quoting Jon Udell

Simon Willison shared a quote from Jon Udell on his blog, which might seem like a casual retweet, but it reflects a growing, deeper anxiety: as AI agents become increasingly capable of writing code, raising PRs, and even deploying, where exactly do developers stand? The seemingly safe term “human in the loop” is quietly handing over agency.

Jon Udell’s criticism is blunt—this phrasing makes humans an accessory to machines. It’s like an automated assembly line where the machine does the work and a human comes by to glance and press a confirmation button. That’s precisely the future we don’t want. He proposes an alternative: agent-assisted process. The difference lies in who owns the loop. In an agent-assisted process, the loop is defined by us developers, and AI agents are invited collaborators—not the other way around, where we are inserted into the AI’s closed loop as rubber stamps.

The implications for AI engineering are far greater than they appear at first glance. If a team always treats AI as a “tool,” they design interactions that place humans in a passive supervisory role, gradually eroding our system understanding and judgment. But once we treat AI as a “team member,” we must more actively define interfaces, collaboration protocols, and review mechanisms. For example, we shouldn’t let an AI dump an entire feature into a single massive PR; we should require it to break down tasks, provide context, and allow incremental intervention—exactly the logic behind Udell’s plea to “not produce unreviewable PRs.”

At a deeper level, this narrative shift reveals a cultural fracture in AI engineering. We used to learn “how to use tools”; now we need to learn “how to lead non-human members.” The new skills required by managers, architects, and developers are not about tuning parameters or writing prompts, but about designing composable, observable, and interruptible collaboration flows. It’s like onboarding a super-intern: you need to set norms, draw boundaries, and establish checkpoints, rather than letting them work in isolation and signing off at the end.

What practical lessons can ordinary developers draw from this? One immediate change is to start using “agent-assisted” instead of “AI-generated” in code reviews and requirement discussions. This term reminds everyone that AI output is produced within our given context, and we retain the right to reject, rewrite, or even restructure its working methods. Another more actionable suggestion is to create a “collaboration contract” for AI agents—explicitly defining the intermediate artifacts, logs, and reasoning they must provide, ensuring every step is explainable and reversible.

From a broader trend perspective, in the coming years, the teams that truly leverage AI well will not be those treating it as black‑box automation, but those that turn “human-AI collaboration” from a slogan into engineering practice. Jon Udell’s words are a signal: it’s time to reclaim ownership of the loop.