evaluator-optimizer

10 hands-on demos covering every major agentic workflow design pattern — built with the OpenAI Agents SDK

Autonomous supervisor → worker → critic loop for Claude Code — delegate one goal, an independent critic grades each round against a rubric, re-dispatch until done. 主Agent监督子Agent+独立评委，定验收标准反复迭代到完美，全程不问你。

Arabic-first multi-agent data-cleaning system - the LLM plans, deterministic pandas executes. Measured quality lift 83 to 97 with human-in-the-loop approval and a full audit log.

A reusable, domain-agnostic self-improving loop for Claude Code — an adversarially-gated evaluator–optimizer that runs unattended without fooling itself or wrecking the machine.

A zero-dependency governance and evidence-integrity kernel for any agent runtime.

A financial terminal interface designed for natural language climate intelligence, bond model interrogation, automated underwriting audits, and portfolio hedging generation.

Multi-agent content pipeline that loops generator, evaluator, and refiner agents until output passes your voice rules. Built with Claude Agent SDK. Forkable for any voice.

Builder/critic iterative improvement loop for Claude Code — evaluator-optimizer skill with fresh independent critics, frozen goal contracts, and audit trails