Apparently there’s an open-source framework, freephdlabor, that aims to make research automation adaptive rather than scripted. A manager process coordinates task-specific workers via files; workers decide whether to dig deeper, pivot, or call others based on results. In a case study on neural-network training phases, it reportedly found phase transitions and designed follow-up tests, producing a paper draft. The pitch is that you can swap in domain-specific workers and keep the same coordination layer.
looks fun, I think the main advantage of it compared alternatives is that it's actually customizable? which could be important as you can't expect current systems to work out-of-the-box on your use cases.
GitHub: https://github.com/ltjed/freephdlabor
Paper: https://arxiv.org/abs/2510.15624
Demo: https://freephdlabor.github.io/