FreePHDLabor
Overview
FreePHDLabor is an open-source multiagent framework, presented in the paper "Build Your Personalized Research Group: A Multiagent Framework for Continual and Interactive Science Automation" (arXiv:2510.15624), that automates the complete scientific research lifecycle — hypothesis generation, experimentation, and publication-ready manuscript writing — with fully dynamic, real-time-reasoning-driven workflows rather than a fixed pipeline. Its stated goal is to let a researcher customize a personalized multiagent "research group" tailored to a specific scientific domain and have it run continually against a research problem, in some cases producing an executed paper with experiments, figures, and citations within hours.
Key Concepts / Architecture
FreePHDLabor uses a hierarchical multiagent design coordinated by a central orchestrator:
- ManagerAgent — the central orchestrator; receives a research task from the user via the command line and delegates work to specialized agents according to the current research phase.
- IdeationAgent — generates and refines research hypotheses.
- ExperimentationAgent — drafts code and executes experiments (incorporates modified components from AI-Scientist-v2).
- WriteupAgent — produces academic papers with LaTeX formatting, figures, and citations.
- ReviewerAgent — provides scholarly feedback on drafts and results.
- ProofreadingAgent — handles formatting and copyediting of the final manuscript.
Agents communicate through a shared workspace directory rather than a rigid message-passing schema, which is what allows the workflow to adapt dynamically: the system re-plans in real time based on experimental findings instead of following a predetermined sequence of steps.
Technology Stack
- Built on the smolagents framework for agent coordination
- Experimentation pipeline incorporates modified components from AI-Scientist-v2
- Phoenix used for telemetry and monitoring
- Supports multiple LLM providers (OpenAI, Anthropic, Google) via a
.llm_config.yamlfile - Python 3.11+, Conda environment management
- HPC/SLURM-cluster compatible; CUDA-compatible GPU support recommended
Key Features
- Dynamic workflows — the research plan adapts in real time to experimental findings rather than executing a predetermined path.
- Full customization — the modular agent design lets users add, remove, or modify agents to fit a domain-specific research workflow without a complete system redesign.
- Human-in-the-loop — human feedback can be injected naturally during research execution.
- Continual research — context management is designed to sustain exploration of a research problem across multiple sessions.
- Interruption and resume — full workspace state is preserved, so a run can be paused and later resumed without losing progress.
Suitable for (Pros)
- Research groups and labs wanting to prototype an automated, domain-specific research assistant without building multiagent orchestration from scratch
- Continual, long-running exploration of a single research problem across sessions (interrupt/resume support)
- Teams that need to customize the agent roster (add/remove agents) rather than use a fixed research pipeline
Limitations (Cons)
- Requires GPU/HPC infrastructure for realistic experimentation workloads (CUDA, SLURM templates)
- Detailed quantitative evaluation metrics are not published in the repository documentation itself — evaluation evidence lives primarily in the accompanying arXiv paper
- Narrow focus on the ML/AI research idiom (built on AI-Scientist-v2 components); applicability to non-computational scientific domains is unproven in the public materials
Licensing
MIT License — open source, permissive for academic and commercial use.
See Also
- Multi-Agent Systems — hierarchical coordination patterns (ManagerAgent + specialized workers) that FreePHDLabor exemplifies
- Self-Learning Agents Reference Architecture — related continual-improvement and automated-research patterns
- Discovery Loop — a commercial platform pursuing large-scale parallel automation of ML research and science
- Agent Development Frameworks — comparison of general-purpose agent frameworks