bytedance/deer-flow
DeerFlow is a community-driven deep research framework that integrates language models with web search, crawling, and Python execution tools to facilitate comprehensive AI-powered research workflows.
Awesome AI Agents › Reasoning Frameworks
Tencent/CognitiveKernel-Pro
Cognitive Kernel-Pro is an open-source framework from Tencent for building deep research agents and for training the foundation models that drive them. The runtime is a hierarchy: a main agent plans and delegates, and specialised sub-agents handle web browsing, file understanding and code execution. The web agent runs against a Playwright-backed browser server that the user hosts locally, and file handling covers PDFs, Office documents, audio transcription and video transcripts through a set of ordinary Python libraries. Generated Python is executed directly, so the documentation instructs users to run the stack in a sandbox with sudo disabled. Models are selected per agent through a configuration dictionary, and any mixture of hosted APIs and locally served vLLM endpoints can be used, which lets a cheap model handle browsing while a stronger one plans. A design goal is to rely on free tools wherever possible: the only paid dependency is the Google Search API, and DuckDuckGo can replace it. The other half of the project is training. The authors release the query dataset, a supervised fine-tuning dataset, and a Qwen3-8B model trained with the published recipe, which they report as outperforming reinforcement-learning-trained web agents such as WebDancer and WebSailor without reinforcement learning; a subgoal-GRPO reinforcement learning recipe is provided for further gains. Related work released alongside it covers synthesising deep research training data, test-time rubric-guided verification for self-evolution, and offline reinforcement learning that assigns credit to critical steps. The audience is researchers and engineers who want a reproducible open stack for deep research agents rather than a hosted product.
https://github.com/Tencent/CognitiveKernel-Pro
Also in Reasoning Frameworks
VectifyAI/PageIndex
PageIndex is a reasoning-based document indexing and retrieval system that uses a hierarchical tree structure to enable human-like, vectorless retrieval and multi-step reasoning over long professional documents.
guidance-ai/guidance
Guidance is a programming paradigm and library that enables efficient and fine-grained control over large language models to produce structured, high-quality outputs with reduced latency and cost.
friuns2/BlackFriday-GPTs-Prompts
A curated repository of free GPT prompts and jailbreaks across various domains, enabling users to utilize GPT models without a subscription to GPT Plus.
refly-ai/refly
Refly is an open-source AI-native creation engine that combines multi-threaded dialogues, multimodal inputs, knowledge base integration, and AI-powered tools to transform ideas into production-ready interactive content and web applications.
mongodb-developer/GenAI-Showcase
MongoDB's GenAI Showcase is a comprehensive repository offering examples and applications that integrate MongoDB with advanced Generative AI techniques like Retrieval-Augmented Generation and AI Agents to support AI development and innovation.
anysphere/priompt
Priompt is a JSX-based prompting library that uses a priority system to efficiently construct token-limited prompts for language models, inspired by React's component design.
HKUDS/Auto-Deep-Research
Auto-Deep-Research is an open-source, fully automated, and cost-effective personal AI assistant that supports multiple large language models and offers a one-click launch experience as an alternative to OpenAI's Deep Research.