Xiangyue-Zhang/auto-deep-researcher-24x7

Imagine queuing up a dozen ablation runs before you go to bed and waking up to results, logs, and a summary of what worked. That is the core promise here. The system uses a Leader-Worker architecture where a coordinating agent breaks down your research goal, dispatches experiment tasks to worker processes, and tracks outcomes without bloating memory over time. The constant-size memory design means it does not collapse under long overnight sessions the way naive LLM loops tend to do. There is no paid monitoring service bolted on, which removes a common friction point for solo researchers watching costs. The honest reservation is that this is genuinely early-stage. You should expect rough edges, limited documentation, and some manual setup before it fits cleanly into your existing training pipeline. It works best if you already have modular experiment scripts you can point it at rather than expecting it to structure your research from scratch. -> Best for: ML researchers and PhD students who run repeated GPU experiments and want a lightweight autonomous loop without subscribing to another monitoring tool.