auto-deep-researcher-24x7

Why we included this project

If training runs keep eating your evenings, this agent is built to take over the watching. It uses a Leader-Worker design where a leader decides what to try next and workers execute individual runs, with a constant-size memory that keeps context manageable over sessions stretching across days. Recent work added a Slurm backend that submits jobs from a login node and reads real terminal states, plus presets for Chinese LLM APIs, so you can drive the whole loop without paying for a Claude or Codex subscription. It suits researchers and small ML teams who want an always-on assistant to push through hyperparameter and architecture sweeps overnight, rather than a general framework for building your own agent pipelines. Expect to wire up a coding-agent CLI, a GPU, and your own experiment brief before it runs.