Human-Supervised Agent Loop Orchestrator

The problem

Uncontrolled agent loops quickly become expensive, hit rate limits, or produce hallucinated results when left without strict boundaries. Developers often struggle to define exact stop criteria and human escalation rules that prevent "infinite loops" and budget drain.

What it does

• Classifies loops into specific architectures like goal-based, time-based, or proactive triage.
• Generates concrete Definition of Done rubrics and judge-based validation layers.
• Hardcodes stop criteria, retry limits, and token budget constraints into the design.
• Maps tool permissions and identifies deterministic script opportunities to replace expensive reasoning steps.
• Produces a standardized loop receipt to track status and human review triggers.

Why this beats prompting it yourself

Prompting a loop manually usually misses edge cases like retry exhaustion or judge-model tiering. This skill enforces a rigorous architectural framework that ensures safety protocols, like escalation rules and budget caps, are never ignored. It shifts the design from vague automation to a reliable, human-supervised systems engineering task.

Use cases

• Designing a security triage loop that flags high-severity vulnerabilities for manual review.
• Building a recurring research agent that stops after reaching a specific token budget.
• Creating a support inbox classifier with a separate judge model to validate accuracy.
• Orchestrating a weekly QA loop that generates a formal loop receipt for engineering leads.

Known limitations

This skill provides the scaffold and plan for a loop but does not execute the code or provide the underlying compute infrastructure. It requires a human to approve high-impact actions like financial transactions or production deployments.