DeepSearchAgents

VIBE 🖖 Build with 💖 for Humanity with AI

<h2>

<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/smolagents/smolagents.png" alt="Smol Pingu" height="25">

<img src="https://raw.githubusercontent.com/modelcontextprotocol/modelcontextprotocol/main/docs/logo/dark.svg" alt="MCP" height="25">

🚅

<img src="public/Jina-white.png" alt="Jina AI" height="25">

</h2>

Come From Open Source, This is the Way

中文版

README Update Date: 2025-08-10

[v0.3.3.dev] Dev Status: "✅ Complete the development and integration of the DSCA front-end Alpha version"

1. Introduction

🚀 CODE ACT IS ALL YOU NEED

The DeepSearchAgent project embodies the philosophy that executable code as action is the most powerful paradigm for AI agents. By treating code generation and execution as the primary means of interaction with the world, we unlock unprecedented flexibility and capability in autonomous systems.

The DeepSearchAgent project is an intelligent agent system based on the ReAct (Reasoning + Acting) reasoning and action framework and the CodeAct ("Code as Action") AI agent concept. It aims to realize broader task reasoning and execution through "DeepResearch" DR-Multi-Agent, leveraging DeepSearch's multi-step network deep search foundational capabilities. It utilizes the reasoning power of AI language models (LLMs), along with a toolbox collection and programming action invocation abilities within a Python packages sandbox, enabling it to handle complex web search tasks that are both broad and deep via multi-step deep search, multimodal webpage text processing, reading, and multi-step reasoning. The system also provides traceable reference materials. Built upon Hugging Face's smolagents framework, this project implements a dual-mode intelligent agent system capable of invoking predefined toolboxes as well as writing action code—realizing both "generation of dedicated dynamic DSLs based on task plans" and "AI self-created dynamic one-time dedicated tools.

The project supports a CLI TUI interface for terminal command-line operation, a standard FastAPI service, the FastMCP MCP server, and a modern WebTUI frontend with terminal-style aesthetics. The WebTUI (v0.3.3) features real-time WebSocket streaming, a cyberpunk-inspired design system, and optimized performance for displaying the CodeAct Agent Run process. It facilitates developers in experimentation, integration, and usage across various systems. This is an open-source project aimed at providing VIBER beginners with a friendly Code Agent experience, learning opportunities, and extensibility.

2. ✨ Features

• 👻 Deep Search Task Capability: Handles complex questions through multi-step searching, reading, and reasoning processes involving online content.
• DeepSearch Specialist: Supports both CodeAct (Python code execution) mode and ReAct (tool invocation) mode for experimental comparison; configuration of related Agent runtime, language models, and various tools can be managed via config.toml (src/core/config/settings.py).
• 🪄 Extensible Toolbox: Built-in set of tools for web searching, content retrieval, text processing, semantic ranking, computation, and GitHub repository analysis.
• 🌐 Hybrid Search Engine (v0.3.1): Multi-provider search aggregation supporting Google (Serper), X.com, Jina AI, and Exa Neural search with intelligent deduplication and ranking.
• 🌐 Web API v2 with Real-time WebSocket Streaming (v0.3.2): Simplified Gradio message pass-through architecture for web frontend integration with real-time agent execution visibility.
• 🔍 Text Embedding and Re-ranking: Uses Jina AI embedding and re-ranking models to process multimodal web content from URLs.
• 📚 GitHub Repository Q&A (v0.3.1): AI-powered repository analysis tool using DeepWiki MCP for understanding GitHub projects.
• 🐦 X.com Deep Query (v0.3.1): Specialized tools for searching, reading, and analyzing X.com (Twitter) content using xAI's Live Search API.
• 🧠 Periodic Planning and Updates: Implements strategic reevaluation during execution to optimize search paths.
• 🔄 Iterative Optimization: The AI specialist continuously improves search and analysis strategies based on initial findings through self-optimization, and continuously optimizes task execution paths by updating the task plan to achieve task objectives.
• 💻 Multiple development debugging and user interaction modes: Provides CLI command-line interaction, standard FastAPI service, and Web GUI frontend service
• 🔗 Traceable References: Provides sources and references for generated answers.
• 📺 Streaming Output: Supports real-time streaming of agent steps and final answers, with rich text formatting.
• 🧮 Symbolic Computation: Integrated WolframAlpha symbolic computation engine, supporting mathematical and computational problems
• 📝 JSON/Markdown Rendering: Automatically detects and presents structured outputs in user-friendly formats.
• 🤝 Hierarchical Multi-Agent Support (v0.2.9): Manager agent mode orchestrates teams of specialized agents for collaborative problem-solving.
• ⚡ Parallel Tool Execution (v0.2.9): Multiple concurrent tool calls for improved performance and efficiency.
• 📊 Enhanced Execution Metrics (v0.2.9): RunResult objects provide detailed execution metadata including token usage and timing.
• 🔒 Improved Security (v0.2.9): Latest security patches from smolagents v1.17.0-v1.19.0 applied.
• 🧠 Structured Generation (v0.2.9): Optional structured outputs for CodeAgent improving reliability.
• 🔄 Context Manager Support (v0.2.9): Proper resource cleanup lifecycle for better memory management.
• 💾 Enhanced Memory Management (v0.2.9): Agent memory reset and summary capabilities for long-running sessions.

Reference Use Cases (To be updated v0.3.1+)

WebTUI Demo:

<video src="public/DSCA-web-demo.mp4" controls="controls" style="max-width: 100%;"> </video>

• CodeAct Mode Example: Full CLI run showing multi-step deep search process.

  • Start:

  

  

  • FinalAnswer:

Development Plan Currently Under Intensive Iteration:

1. [DONE] Developed Web API v2 with real-time WebSocket streaming (v0.3.2) - Simplified Gradio message pass-through architecture replacing complex event-driven system (~5000 lines reduced to ~500 lines). Frontend development and Docker packaging pending;

2. [DONE] Added MCP Client/MCP tools HUB to DeepSearchAgents' DeepSearchToolbox, supporting MCP Tools configuration and invocation;

3. [DONE] Provided packaging of DeepSearchAgents as an MCP server, offering DeepSearchAgent MCP tools services;

4. [DONE] Supported multi-vertical search engine source aggregation (Google, X.com, Jina AI, Exa Neural) with hybrid search aggregation and intelligent result deduplication (v0.3.1);

5. [DONE] Upgraded to smolagents v1.19.0 with hierarchical agent management, parallel tool execution, and enhanced streaming architecture;

6. [DONE] Add a DeepWiki Remote MCP tool to enhance the GitHub URLs vertical crawler/parser with GitHub Repository Q&A capabilities (v0.3.1);

7. [Partially supported in the tool layer] The deep search strategy provides more strategy parameters and adds support for strategy parameters based on Tokens budget;

8. [Experimental version testing] Implement auxiliary methods and tools for Agent Action search width & depth based on Monte Carlo Tree Search strategies in DeepSearchAgents, along with strategy control parameters;

9. [TODO] LLM as Judge: Experimentally add an Agent Runs evaluator for DeepSearchAgents (independently evaluate the deep search paths & results of DeepSearchAgents);

10. [TODO] Add persistent memory layer functionality for Agents & provide users with persistent search records;

11. Add suitable open-source sandbox (E2B-like) adapted code_sandbox Docker automation configuration, and increase support for more remote code_sandbox secure environment SDKs;

12. Integrate full-process Agent Runs telemetry adaptation ("OpenTelemetry" & Langfuse) (integrated together with the Docker packaged version);

13. [TODO] Human-in-the-loop & multi-path branching backtracking functionality for Agent Runs;

14. [Experimental] Special implementation version of multi_agent_HiRA (Hierarchical Reasoning Framework for Deep Search) based on special tokens protocol (arXiv-2507.02652v1);

15. [Experimental] Add auxiliary method optimization to the agent omni-tools-query pipeline based on [submodular-optimization] ("submodular optimization algo