# Deep Researcher Agent

An autonomous agent that runs deep-learning experiments around the clock on a local machine, a remote server over SSH, or a Slurm cluster, tracking real job outcomes with constant-size memory and zero-cost monitoring.

Deep Researcher Agent autonomously runs a user's deep-learning experiments continuously, using a leader-worker architecture and constant-size memory rather than an ever-growing context, with monitoring that costs nothing while jobs run. It can drive experiments locally, over SSH, or on a Slurm cluster, where training is submitted with `sbatch --parsable` over a single transient SSH call so no process is left running on the login node, and job liveness is read from `sacct` and `squeue` with bounded checks that guarantee the monitor loop eventually terminates.

The agent records a job's true terminal state, such as failed, timed out, or cancelled, rather than always marking a finished job as completed, and reasons over that outcome when deciding what to try next. It can run against a Chinese LLM API instead of a Claude or Codex subscription by setting a one-word provider preset (deepseek, qwen, kimi, or glm) that fills in the matching OpenAI-compatible base URL and key environment variable, with the same path also usable for self-hosted or proxied endpoints.

It is written in Python and Apache-2.0 licensed.

Website: https://arxiv.org/abs/2604.05854