MetaGPT is a multi-agent orchestration framework designed to automate software engineering workflows and complex task execution. It functions as a Python library and CLI tool where autonomous agents (Roles like Product Manager, Architect, Engineer) collaborate via a message-passing environment to generate code, designs, and documentation. The system abstracts LLM interactions, supports Retrieval-Augmented Generation (RAG) for knowledge retrieval, and includes specialized environments for simulation (Minecraft, Stanford Town) and automation (Android, Web Browsing). It is primarily an application framework rather than a data research repository, though it incorporates research-grade components like Tree of Thought planning and AutoML evaluation. The architecture is modular, separating Roles, Actions, Tools, and Environments to allow for extensibility. Key entry points include the CLI command `metagpt` and the `metagpt.software_company` module for project generation. The codebase relies heavily on external dependencies, particularly LLM APIs (OpenAI, Azure, etc.) and browser automation tools (Playwright), which introduces maintenance complexity. While the core framework is well-structured with factory patterns and serialization support, the heavy coupling to external services and the presence of complex, niche extensions (e.g., Minecraft bot integration) contribute to a medium technical drift rating.
QwenLM/Qwen-AgentWorld is a research benchmark evaluation toolkit for Language World Models (LWMs). The repository contains the AgentWorldBench evaluation infrastructure: a standalone CLI script (`eval/eval.py`) that orchestrates a two-stage evaluation pipeline — first calling an OpenAI-compatible inference endpoint to generate environment predictions from a world model, then using an LLM judge (e.g., GPT) to score the predictions across five dimensions (format, factuality, consistency, realism, quality). The code covers seven agent domains (MCP, Search, Terminal, SWE, Android, Web, OS) through per-domain prompt templates and task configurations. The repository is not a model serving system or training framework; it is a reproducible evaluation harness published alongside the Qwen-AgentWorld model weights and benchmark dataset (hosted separately on Hugging Face). The core value is in the robust parsing utilities for extracting structured output from LLM responses, including sophisticated handling of think/response tag removal and multi-strategy JSON extraction from judge outputs.