Loop engineering: design the system that prompts the agent
Jun 8 Addy Osmani: automations, worktrees, skills, connectors, maker/checker sub-agents, plus disk memory — /goal stops on a separate model’s done-check, not the implementer’s optimism.
Source: Loop Engineering
Addy Osmani’s Jun 8 essay names the floor above the harness: loop engineering is replacing yourself as the person who prompts the agent. Steinberger and Cherny’s line — design loops that prompt Claude/Codex — becomes a portable anatomy once both products ship the same five pieces.
- Automations as heartbeat (Codex Automations / Claude
/loop, cron, hooks, Actions) so discovery and triage fire without a human in the chair. - Worktrees so parallel agents cannot collide on the same checkout.
- Skills so project intent compounds on disk instead of being re-guessed every session.
- Plugins/connectors (MCP) so the loop opens PRs and updates tickets instead of narrating what it would do.
- Sub-agents that keep the maker away from the checker — explorer / implementer / verifier, sometimes different models.
The sixth piece is boring and load-bearing: memory outside the conversation (markdown, Linear, anything durable). Models forget between runs; the repo does not. The in-session primitive that matches the thesis is /goal in both tools: keep going until a verifiable condition holds, with a fresh small model grading done so the writer is not the stop condition. Automations surface work; worktrees isolate it; skills and connectors make it act in your real environment; sub-agents and /goal make “done” mean something you can walk away from.
Osmani’s caution is the productively complicating half. Loops sharpen three debts: unattended mistakes, comprehension debt as unread code grows, and cognitive surrender — taking whatever the loop ships because it is comfortable. Two people can build the same loop; one uses it to move faster on work they understand, the other to avoid understanding. The loop cannot tell which. Direct prompting still works; the leverage point moved, the engineer did not disappear.
Why it matters for code review
Continuous overnight momentum is this stack — schedule, isolate, verify with an independent checker, persist state — not a longer peak Claude Code session. Review bandwidth remains the ceiling on how many worktrees you can actually run; FOMO about idle agents is the wrong scarcity if the loop has no external done-check or the human has stopped reading what it merges.