DeepResearchAgent is a hierarchical multi-agent system designed for “deep research” tasks and general-purpose problem solving. The framework decomposes high-level tasks via a top-level planning agent and coordinates multiple specialized lower-level agents (e.g., deep analyzer, deep researcher, browser automation agent, MCP manager) to collaboratively execute complex workflows. It supports asynchronous operations and both local and remote model inference.
Key Features
- Hierarchical planning: a top-level planner decomposes and schedules tasks while lower-level agents handle fine-grained execution and tool calls.
- Rich tool integrations: supports browser automation (for information collection), a secure Python execution sandbox, and MCP-based tool registration and invocation.
- Multi-model support: compatible with OpenAI, Anthropic, Google LLMs, and local vLLM (Qwen series) backends.
- Multimedia capabilities: includes image and video generation tools (Imagen, Veo3) for multimodal experiments and demos.
Use Cases
- Automated literature search and summarization: search, read, and summarize academic papers or web resources.
- Complex task orchestration: suitable for scenarios requiring decomposition of high-level goals into multiple asynchronous subtasks (e.g., large-scale data analysis, experiment design).
- Web data collection and interaction: paired with browser agents for navigation, scraping, and automated interactions for data gathering and verification.
Technical Highlights
- Asynchronous architecture: built on async programming to improve concurrent task handling.
- Extensible agent system: support for adding custom specialized agents to extend capabilities.
- Secure sandboxing: restricted Python tool execution (import controls, resource limits) to reduce RCE risk.