import { ArticleLayout } from '@/layouts/article-layout' export default ArticleLayout

Loop Engineering: A Deep Retrospective on a Field Failure

When the Design Documents Cannot Be Executed

On June 7, 2026, OpenClaw founder Peter Steinberger posted: "You shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents." The same day, Anthropic Claude Code lead Boris Cherny said publicly: "I don't prompt Claude anymore. I have loops running. My job is to write loops."

On June 8, Google engineering director Addy Osmani published a long article formally naming this methodology Loop Engineering.

On June 13, we spent an entire day trying to build a complete system from scratch.

Result: failure.

---

What We Actually Built

36 files. 7 modules. 10 cron jobs. 8 Loops. 1 Dashboard.

On paper, this is a complete Loop Engineering system:

• A three-layer architecture (Main → Looper → coder-*)
• A seven-phase inner loop (Observe → Plan → Execute → Verify → Diagnose → Adjust → Retry)
• A cap of 30 self-corrections + a four-level escalating strategy
• A five-dimensional meaning evaluation model (only ≥70 points gets converted into a Loop)

The Truth: Nothing Happened

When we used openclaw cron run to trigger a deliberately failing test Loop, the file was not created, and STATE.json was completely empty.

Root cause: there was no Loop Runtime Engine.

The YAML frontmatter of LOOP.md defined executor, checker, inner_loop, and escalation. But after cron triggered, there was no code at all to parse and execute those definitions.

Here is the gap between industry standards and our reality:

Industry Standard	Us
Maker ≠ Checker	❌ At the design stage, both were the looper
Separate verifier model	❌ Written, but no code executed it
/goal: run until condition met	❌ The inner loop never happened
State on disk, not in context	✅ Designed correctly

Three Fatal Mistakes

1. A file ≠ a system. 36 design documents are not the same as a runnable system.

2. sessions_spawn ≠ a Loop. We used sessions_spawn to bypass the inner loop and complete all our "deliverables". This is not Loop Engineering; it is scheduling multiple Agents.

3. Blind spots are only exposed at runtime. The Maker/Checker Split error was completely undetected at the design stage, because the system had never been executed.

Next Steps

Core insight: the hard part of Loop Engineering was never writing LOOP.md, but writing the engine that can execute LOOP.md.

Three paths: a Python Loop Runner (a 200-line build of our own), Claude Code /goal (a native mechanism), or a hybrid approach.