Jon Udell wants you to drop the phrase “human in the loop” entirely. His argument: the phrase cedes authority to the machine. It frames the process as belonging to the AI, with you occasionally invited to rubber-stamp outputs.
Udell’s reframe is specific:
- It’s your loop. You work the way you always have — planning, reviewing, deciding. Agents get recruited onto the team, not the other way around.
- Agent-assisted work isn’t a black box. A process that “takes in prompts and emits features” is one you’ve already lost control of.
- Unreviewable PRs are a choice. Udell’s punchline lands like a doctor joke: “Doctor, it hurts when agents create unreviewable PRs.” “Don’t do that.”
The distinction matters because the default agentic workflow at most companies right now looks like this: point an agent at a ticket, let it generate a branch, review whatever comes back. Udell is saying that’s backwards. You don’t adapt your review process to fit the agent’s output size — you constrain the agent’s scope so its output stays reviewable.
This isn’t abstract philosophy. Teams shipping with Claude Code, Cursor agents, or Codex are already hitting the wall where a single agent run touches 15 files across three modules and the resulting PR is effectively unauditable. The “human in the loop” framing encourages you to accept that — you’re “in the loop,” after all, you approved the merge. Udell’s framing forces a harder question: did you actually understand what changed?
NOTE
Udell’s post was surfaced by Simon Willison, who tagged it under coding-agents and agentic-engineering — two categories where the “who owns the process” question keeps recurring.
The practical takeaway is small enough to act on today: if your agent-generated PRs are too large to review line-by-line, the fix isn’t better review tools — it’s smaller agent tasks. Keep the loop yours.