evaluator-optimizer

两仪 Liangyi · 跨厂商多 Agent 产品想法优化工作流

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。

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

Agent SDK tool that runs your project's Sprint Zero — asks the product-shape questions, inspects the codebase, and writes AGENTS.md + context/ so every new project starts with a real operating manual instead of a blank prompt. Uses the evaluator-optimizer pattern (a subagent verifies the output before reporting) and ships as a portable SDK plugin

Self-checking market-research pipeline: web-grounded drafting, per-section critic loop with model-chosen re-search, and a ship/hold quality gate (exit code 3) · Конвейер рыночных отчётов с циклом критика по разделам и гейтом качества

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.

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.

Runnable implementations and documented failure modes for the Generator-Critic (Reflection / Evaluator-Optimizer) LLM pattern

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