The Agency (agency-agents) is an open-source project by msitarzewski, offering a collection of meticulously crafted AI agent personalities designed to optimize workflows as specialized AI experts. Each agent possesses deep domain expertise (e.g., frontend development, backend architecture, UX research, PPC strategy), a unique personality, defined processes, and measurable deliverables. Far from generic prompt templates, these are production-ready solutions that integrate seamlessly with various development tools like Claude Code, Gemini CLI, and OpenClaw, enabling users to efficiently automate tasks as if having a dedicated, always-on AI specialist team.
Paper2Code is a tool powered by PaperCoder, a multi-agent Large Language Model (LLM) system, designed to automate the generation of executable code repositories directly from machine learning scientific papers. It employs a three-stage pipeline—planning, analysis, and code generation—each managed by specialized agents. This method has shown superior performance on Paper2Code and PaperBench benchmarks, producing faithful and high-quality implementations, supporting both OpenAI API and open-source models via vLLM.
AutoResearchClaw is an autonomous research agent framework within the OpenClaw ecosystem that transforms research ideas into conference-ready papers. It features an end-to-end 23-stage pipeline covering literature review, hardware-aware sandbox experiments, and LaTeX drafting. Key capabilities include a Human-in-the-Loop (HITL) Co-Pilot system with 6 intervention modes, anti-hallucination citation verification, and self-evolving learning via MetaClaw integration. It supports various LLM backends and integrates with messaging platforms through OpenClaw.
Meta-Harness is a framework developed by Stanford IRIS Lab for automated search and end-to-end optimization of task-specific model harnesses—the code surrounding a base model that manages storage, retrieval, and context. It automates the discovery of optimal scaffolding and memory systems using proposer agents (like Claude Code) to iteratively refine harness code, as demonstrated in text classification and Terminal-Bench 2.0 experiments.
