How does the skill prevent performance regressions?

It uses strict regression boundaries defined in the journal; if any critical metric drops or a guardrail is violated, the system enforces an immediate rollback.

What is the 5-phase execution loop?

The loop consists of Init (context building), Hypothesize (proposing changes), Experiment (applying changes), Analyze (evaluating results), and Compound (deciding to keep or rollback).

Do I need to manually run each phase of the optimization?

No, the skill is designed to auto-advance through phases. It will only pause for user confirmation at key decision gates (Hypothesize and Analyze) before continuing.

What are 'levers' in this optimization skill?

Levers are adjustable parts of your agent, such as prompt templates, system configurations, grader logic, or specific code blocks that affect the agent's output.

Where is the optimization state stored?

All state, including history and metadata, is stored in a local journal.yaml file located within the .claude/optimization-loops/ directory.

Agentic Optimization Controller

Name: Agentic Optimization Controller
Author: mberto10

bymberto10

•

分析と監視

Orchestrates a structured five-phase loop to systematically improve AI agent performance through data-driven experimentation and regression testing.

The Agentic Optimization Controller is a specialized skill designed for developers who need a rigorous, repeatable process for fine-tuning AI agents. It implements a strict 5-phase protocol—Init, Hypothesize, Experiment, Analyze, and Compound—to iterate on 'levers' such as prompts, configurations, and code. By integrating deeply with langfuse-analyzer and maintaining a persistent state journal, the skill ensures that every change is validated against a baseline, preventing regressions while identifying the most effective paths for performance gains.

主な機能

01Persistent state journaling to track metrics, targets, and iteration history

021 GitHub stars

03Automated 5-phase execution loop for systematic agent improvement

04Strict regression guardrails that automatically trigger rollbacks on failure

05Preflight contract validation to ensure data and dataset integrity

06Single and multi-lever tuning strategies for granular or holistic optimization

ユースケース

01Refining complex LLM prompts by testing hypotheses against historical benchmarks

02Continuous performance monitoring and automated regression testing for production agents

03Optimizing agentic workflows by tuning hyperparameters and configuration settings

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add mberto10/mberto-compound optimization-controller

For use in Claude.ai and ChatGPT

主な機能

01Persistent state journaling to track metrics, targets, and iteration history

021 GitHub stars

03Automated 5-phase execution loop for systematic agent improvement

04Strict regression guardrails that automatically trigger rollbacks on failure

05Preflight contract validation to ensure data and dataset integrity

06Single and multi-lever tuning strategies for granular or holistic optimization

ユースケース

01Refining complex LLM prompts by testing hypotheses against historical benchmarks

02Continuous performance monitoring and automated regression testing for production agents

03Optimizing agentic workflows by tuning hyperparameters and configuration settings

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add mberto10/mberto-compound optimization-controller

For use in Claude.ai and ChatGPT