Freephdlabor: Customizable multiagent research automation system (freephdlabor.github.io)

freephdlabor is an open-source, fully customizable multiagent framework that automates the entire research lifecycle—from idea generation to LaTeX-formatted papers with figures and citations—aiming to give researchers “24/7 PhD labor.” Unlike prior fixed‑pipeline systems, freephdlabor uses dynamic workflows and a hub‑and‑spoke ManagerAgent architecture so you can spin up a tailored multiagent lab for your domain within hours. The project targets a practical step toward accelerating scientific progress by enabling continuous hypothesis testing, experiment runs, and iterative writeups with human-in-the-loop checkpoints. Key technical innovations address three core challenges: workflow flexibility, context-window limits, and multiagent coordination. A single ManagerAgent maintains a condensed research history and agent descriptions and delegates tasks to specialized agents (Ideation, Experimentation, ResourcePreparation, Writeup, Reviewer) via tool-like calls and a final_answer protocol. Context is offloaded to a shared workspace of descriptive files (avoiding token bloat and “telephone” errors), while AutomaticContextCompactor summarizes long conversations when token usage exceeds ~75% of the model limit. The system also supports memory persistence, step replay, secure code execution, and per-agent tool sets to avoid quadratic context growth and improve debuggability. The result is a modular, maintainable platform for domain-specific, long-running autonomous research workflows.