Deep Researcher Agent

Deep Researcher Agent is an open-source framework by Xiangyue Zhang that lets LLM agents run deep learning experiments continuously. It covers editing code, launching training, monitoring runs, parsing logs and choosing the next experiment, driven by a project brief written by the researcher. It is released under the Apache 2.0 license and described in an arXiv technical report (2604.05854).

What It Is

The framework is an experiment-operations agent for deep learning researchers who already know what they want to test. It runs a THINK, EXECUTE, REFLECT loop: it reads PROJECT_BRIEF.md, plans the next experiment, modifies code or configs, performs a dry-run, launches GPU training, then parses results and decides what to try next. A leader agent dispatches one worker at a time (idea, code or writing agent), each with a minimal set of 3-5 tools.

How the Loop Stays Cheap and Bounded

The paper and README describe zero-cost monitoring: during training the agent relies on process checks, nvidia-smi and log reads rather than LLM calls, so the model is only invoked for planning and reflection. A two-tier memory (the project brief plus a rolling MEMORY_LOG.md) is capped at roughly 5K characters regardless of runtime. The authors report 500+ experiment cycles across four projects, a 52% improvement over baseline in one project, and about $0.08 of LLM cost per 24-hour cycle. These figures are the authors' own.

Control and Visibility

Researchers steer the agent through PROJECT_BRIEF.md for stable goals and constraints and HUMAN_DIRECTIVE.md for temporary redirects. Progress is visible through /experiment-status and /progress-report, with optional Obsidian sync or local text notes. Later updates add an experiment ledger, dead-end and insight journals, a stagnation signal, code search and arXiv literature tools, and an optional cycles-per-hour limit.

Setup and Execution Backends

Installation is by cloning the repo, installing Python requirements and running install.py, which installs Claude Code slash commands and Codex local skills. It needs Python 3.10+, at least one NVIDIA GPU and access to an Anthropic-compatible or OpenAI-compatible API, or a Claude/Codex CLI subscription. Execution can be local, over SSH to a remote host, or on a Slurm cluster. Provider presets cover DeepSeek, Qwen, Kimi and GLM.