Loop Architect

The Problem

You want to run an autonomous agent loop — but /goal takes whatever condition you type, however vague. No coaching, no second opinion, no structure. The model grades its own homework. When the loop fails or burns tokens, you have nothing portable to inspect, version, or reuse.

The Solution

Loop Architect is a design coach that helps you scaffold a well-designed agent loop before any runner starts changing files. It interviews you, critiques your design against built-in rubrics, wires in a cross-model reviewer, and emits portable artifacts.

Based on Looper by Kevin Simback, MIT License. Adapted for Kiro CLI with native /goal, subagent, and review loop integration.

When to Use

• Design an agent loop with explicit verification before running it
• Set up a self-review or LLM-as-judge loop with a different model as reviewer
• Build a multi-model council for plan and delivery gates
• Scaffold a /goal-driven process with typed verification and termination guards
• Create portable loop specs that work across CLI tools

Installation

npx skills add https://github.com/fabricioctelles/skills -s loop-architect

How It Works

+--------------------------------+
| 1. Goal + context              |
|    read sources                |
+--------------------------------+
              |
              v
+--------------------------------+
| 2. Draft plan.md               |
|    state -> state.json         |
+--------------------------------+
              |
              v
+--------------------------------+
| 3. Plan gate                   |
|    verdict: reviewer-1         |
+--------------------------------+
  | needs work -> revise <= 3 -> step 2
  | pass
              v
+--------------------------------+
| 4. Write delivery-N.md         |
|    log -> run-log.md           |
+--------------------------------+
              |
              v
+--------------------------------+
| 5. Delivery gate               |
|    verdict: reviewer-1         |
+--------------------------------+
  | needs work -> revise <= 3 -> step 4
  | pass
              v
+--------------------------------+
| 6. Final output                |
|    all gates clean             |
+--------------------------------+

Stops: pass gates | max iterations | no progress x2 | budget cap

The seven-stage interview

1. Goal — coached toward outcome, scope, context, and done state
2. Verification — typed as programmatic, judge, or human
3. Host model — which CLI drives the loop
4. Council — reviewer/judge from a different model family
5. Gates & control — revision caps, no-progress detection, budget
6. Confirmation — ASCII flow preview
7. Emit/run — write artifacts and optionally execute

Three execution paths

Path	When	How
/goal	Simple loop, no cross-model review	/goal --max 12 <definition_of_done>
Subagent pipeline	Cross-model review needed	Native Kiro review loops with NEEDS_CHANGES trigger
Python runner	External, scheduled, or CI	python3 ./loop-architect-output/run-loop.py

Emitted Artifacts

File	Purpose
loop.yaml	Human-authored loop spec
loop.resolved.json	Compiled, validated — the runner reads this
LOOP.md	Human-readable summary + ASCII flow
RUN_IN_SESSION.md	Handoff prompt for in-session execution
run-loop.py	Portable Python runner
loop-workspace/	Working directory with state and logs

Model Detection

Loop Architect detects installed CLIs and their capabilities:

python3 ~/.kiro/skills/loop-architect/scripts/looper.py detect-models

Kiro CLI is detected with full capabilities: headless, goal, subagent, review_loop.

vs /goal and Ralph Loop

	/goal	Ralph Loop	Loop Architect
Layer	execution	execution (spec tasks)	design (pre-flight)
Coaches goal	no	no	yes
Typed verification	no	self-correction	programmatic / judge / human
Cross-model review	no	no	yes, by default
Portable artifact	no	no	loop.yaml + resolved spec

📄 Full documentation on GitHub