MOC: Loop Engineering

Map of Content for the practice of loop engineering — designing systems that prompt your coding agents on a loop, instead of prompting them turn-by-turn.

Overview

Loop engineering is the discipline of building a small system that finds work, hands it out, checks it, records what's done, and decides the next thing — so the system pokes the agents rather than you doing it one turn at a time. The term was coined by Addy Osmani (June 2026), building on ideas from Boris Cherny (Claude Code) and Peter Steinberger (OpenClaw). It sits "one floor above" the agent harness: the harness is what a single agent runs inside; a loop runs on a timer, spawns helpers, and feeds itself.

The five primitives (+ memory)

1. Automations — scheduled discovery/triage.
2. Worktrees — isolate parallel agents from each other.
3. Skills — codify project knowledge the agent would otherwise guess.
4. Plugins/connectors — plug the agent into your existing tools.
5. Sub-agents — one ideates, a different one verifies.
6. Memory — a markdown file / Linear board outside the conversation that holds what's done and what's next ("the agent forgets, the repo doesn't").

Caveats that stay on you

• Verification is still yours — a loop makes mistakes unattended; "done" is a claim, not a proof.
• Comprehension debt — fast loops widen the gap between what exists and what you understand.
• Cognitive surrender — designing the loop is the cure when done with judgement, the accelerant when done to avoid thinking.

Content in This Repository

Essays & Overviews

• loop-engineering - Addy Osmani's definitional essay: the five primitives, the Codex-vs-Claude-Code capability table, /loop and /goal, and why loop design is harder than prompt engineering.
• brittany-ellich-loop-engineering - Brittany Ellich's real-world case study: 108 PRs in ~8 days via a protocol / loop / worker design with a markdown task-board, self-stopping /loop, and capped memory.

Adjacent Practices & People

• boris-cherny-claude-code-tips - Claude Code head on writing loops that prompt Claude.
• trq212-claude-5-context-engineering - Context engineering for Claude 5; relevant to how little explicit instruction modern loops need.
• skill-vs-agents-vs-claude-md - Skills, agents, and CLAUDE.md as loop building blocks.

Key Themes

1. From prompting to designing

The leverage point moved from writing the perfect prompt to designing the system that writes prompts. Same action (using agents), opposite posture (you stay the engineer).

2. Isolation & single-writers

Worktrees keep parallel agents from colliding; making the loop the only writer to the board/memory file prevents corruption when several workers run at once.

3. Self-terminating loops

A loop that can declare itself done (explicit stop condition, bare /loop with no interval) saves tokens and respects the human as the bottleneck.

4. The human moves to the ends

Effort shifts from reviewing code line-by-line to reviewing specs (what goes in the queue) and verifying outcomes (what comes out).

External Resources

• Addy Osmani — Loop Engineering
• Brittany Ellich — 108 PRs in Eight Days · loop-board (open source)
• Peter Steinberger — "You should be designing loops that prompt your agents"
• Boris Cherny — "My job is to write loops"

---

This MOC will grow as we document more loop-engineering practices, tools, and case studies.