freephdlabor Released: An Autonomous Research Framework Based on Dynamic Routing

The open-source multi-agent framework freephdlabor, developed by Ed Li and 6 others, has been released. This system departs from the fixed 'idea generation-experiment-paper writing' assembly line approach of existing research automation agents, adopting a dynamic routing structure where ManagerAgent determines the next step in real time based on research history and agent descriptions.

Dynamic Routing and Hub-and-Spoke Structure

While existing systems (such as AI Scientist-v2) require a complete redesign when changing fields, freephdlabor can adapt simply by swapping tools. It uses a hub-and-spoke structure centered on ManagerAgent without direct connections between agents, preventing context explosion as the number of agents increases. Key agents include IdeationAgent, ExperimentationAgent, WriteupAgent, and ReviewerAgent.

File-Based Communication and Context Management

To prevent information distortion caused by string passing between agents, important information is saved as files in a shared working folder, and only paths and summaries are passed to other agents. Additionally, when token usage exceeds 75% of the model limit, AutomaticContextCompactor activates to summarize and restructure conversation history, preventing context loss during long-term autonomous research.

Validation and Limitations

The authors claim that one paper written by freephdlabor was accepted at the ICLR 2025 workshop, proving it is an 'autonomous research system that has undergone peer review.' However, they explicitly state that the outputs are draft-level premised on human review, and that a fully autonomous system applicable to all scientific fields is not yet possible. Execution requires Python 3.11 or higher, Conda, a CUDA-supported GPU, and API keys from multiple LLM providers.