"You should not be prompting coding agents anymore — you should be designing loops that prompt your agents." — Peter Steinberger
"My job is to write loops, not individual prompts." — Boris Cherny, Anthropic (Claude Code lead)
Loop engineering emerged in June 2026, coined by Addy Osmani (Google), as the next evolution past prompt engineering and context engineering.
The idea: instead of typing each prompt yourself, you build a system that discovers work, prompts the agent, verifies results, persists state, and decides next steps — autonomously, until a goal is met.
| Layer | What you optimize | Who prompts |
|---|---|---|
| Prompt engineering | How you phrase one instruction | You, manually |
| Context engineering | What goes in the context window | You, curated |
| Loop engineering | The system that decides what to prompt, when, and whether the result is valid | The system itself |
Every agent already runs an inner loop internally:
Reason → Act → Observe → Repeat
Loop engineering adds the outer loop:
Discover work → Assign to agent → Verify result → Persist state → Next task → Repeat
This skill implements the outer loop for you.
Any complex goal is iterative. Single-shot generation fails on unexpected conditions, environment-specific issues, and unverifiable output. The ReAct pattern (Reason + Act, from Princeton/Google research) is what makes agents actually converge — they act, observe output, reason about failures, and revise. A loop harnesses this at scale across any domain: code, research, content, data, or automation.
Long agent sessions degrade. The context window fills with dead ends and stale state. The fix: reset context on every iteration and read current state from disk. Each agent turn starts fresh; intelligence lives in clear specs and files on disk, not in a single long session.
This skill implements Ralph automatically — every agent starts with a clean context and reads loop-stack/STATUS.md + loop-stack/PLAN.md for ground truth.
The biggest loop engineering insight: the agent that does the work should not verify its own work. Self-grading is how loops hallucinate progress.
This skill enforces the split: the executor agent does the work, the verifier agent runs the stop condition. The verifier is explicitly forbidden from producing goal output.
| Source | Link |
|---|---|
| Kilo.ai — What Is Loop Engineering? | https://kilo.ai/articles/what-is-loop-engineering |
| Lushbinary — Full Guide (5 building blocks) | https://lushbinary.com/blog/loop-engineering-ai-coding-agents-guide/ |
| MindStudio — ReAct pattern & loop anatomy | https://www.mindstudio.ai/blog/what-is-loop-engineering-ai-coding-agents |
| Cobus Greyling — Loop Engineering overview | https://cobusgreyling.medium.com/loop-engineering-62926dd6991c |
| 36kr — Critical analysis & economics | https://eu.36kr.com/en/p/3864390159366791 |
Key concepts implemented:
- ReAct pattern (Princeton/Google) — the inner loop that makes agents converge
- Ralph technique (Geoffrey Huntley) — fresh context per iteration, state on disk
- Maker/checker split — separate agents for doing and verifying
- Verifiable stop conditions — no vague goals, only testable contracts
- Failure checkpointing — human in the loop only when the loop is stuck