Loop Engineering

Practitioner essay naming and popularizing the layer this library has so far met from the research side: stop prompting agents and design the system that prompts them. Osmani anchors the shift in two quotes — Peter Steinberger (“designing loops that prompt your agents”) and Boris Cherny of Claude Code (“my job is to write loops”) — and positions loop engineering “one floor above” his earlier harness-engineering framing: the harness on a timer, spawning helpers, feeding itself.

The five primitives plus state

The post’s core claim is convergence: what required a bespoke pile of bash a year earlier now ships as first-class features in both the Codex app and Claude Code, under different names but with identical capability, so a loop design is tool-portable. The inventory:

1. Automations — the heartbeat: scheduled discovery and triage, findings landing in an inbox, empty runs self-archiving. Includes an in-session variant the post flags as closest to its thesis: /goal runs until a written stopping condition holds, graded after each turn by a separate model — the maker/checker split applied to the stop condition itself.
2. Worktrees — mechanical isolation for parallel agents; the familiar two-engineers-same-lines collision, solved by git rather than coordination.
3. Skills — project knowledge written down once where every run reads it; the fix for agents that “start every session cold” and fill intent gaps with confident guesses. Without skills the loop re-derives the project from zero every cycle.
4. Plugins/connectors (MCP) — the difference between an agent that says “here is the fix” and a loop that opens the PR and updates the ticket.
5. Sub-agents — “keep the maker away from the checker”: the model that wrote the code is too generous grading its own homework, so a verifier with different instructions (and sometimes a different model) is what makes walking away defensible.

The sixth piece is state outside the context window — a markdown file or issue board holding what’s done and what’s next, because “the agent forgets, the repo doesnt.” The worked example composes all six: a morning automation calls a triage skill over CI failures and issues, writes findings to the state file, opens a worktree per finding with a drafting sub-agent and a reviewing sub-agent, connectors open the PR, and anything unhandled lands in the human’s triage inbox; the state file lets tomorrow’s run resume where today’s stopped.

The counterweights

The closing third argues three problems sharpen as the loop improves: verification stays on the human (“done” is a claim, not a proof); comprehension debt grows faster the smoother the loop ships code you did not write; and the comfortable posture — taking whatever the loop returns — is what he calls cognitive surrender. The line worth keeping: two people can build the same loop and get opposite results — one moves faster on work they understand deeply, the other avoids understanding at all, and the loop cannot tell the difference.

Assessment

This is synthesis and advocacy, not a system report: no measurements, and nearly every load-bearing concept cites one of Osmani’s own prior posts (eleven self-citations — harness engineering, intent debt, comprehension debt, cognitive surrender, orchestration tax, and more), so the essay is best read as the index volume of a personal series. The product-mechanism claims (Codex’s Automations tab and TOML sub-agents, /goal’s separate grading model) are asserted without sources and are the fastest-aging content in the piece. The author has a stake: an O’Reilly book on AI-assisted engineering is promoted in-page, and the coinage visibly became content-marketing fodder within weeks — the ingestion-time Reddit sweep found courses, newsletters, and copycat posts reusing the term, evidence of reach rather than validity. The HN thread is a single sustained objection the post itself concedes: whether comprehension of a production system survives this working style at all.

• Durable: the maker/checker split extended to the stop condition; state on disk rather than in context; loop designs made portable by primitive convergence; the loop-amplifies-the-operator closing.
• Era-bound: the entire product-feature table, already version-specific at publication.
• In this library: names for practitioners the layer weng2026a maps as research territory — Weng’s design patterns (file-system memory, explicit sub-agents, skills) are Osmani’s primitives seen from the product side, and both posts reach the same rule from opposite directions: the grader must sit outside the loop that does the work. Against vincent2026a‘s autonomy ladder, the loop sketched here is delegation without self-improvement — the loop runs the work but nothing updates the loop — which is exactly the gap the Self-evolving agents shelf is about. The external-state principle is karpathy2026‘s compile-don’t-re-retrieve argument applied to task state rather than knowledge.