Deep Researcher Agent: 24/7 Autonomous Deep Learning Experiment Agent

📋 Overview

Deep learning research is often bottlenecked by slow experiment iteration: a researcher designs a run, waits hours for training, checks results, adjusts settings, and repeats this loop across days or weeks. Deep Researcher Agent turns that workflow into an autonomous system. It executes a persistent THINK → EXECUTE → REFLECT loop, where the agent analyzes the project state, decides the next experiment, launches it, tracks the run with process and log monitoring, and then updates its memory before continuing. The design focuses on practical deployment rather than toy agent demos: fixed-size memory for long-running operation, minimal tool schemas, and zero-LLM monitoring while jobs are still training.

📊 Reported Results

🏗️ Core Design

The system combines a persistent leader, lightweight workers, fixed-size memory, and training-time zero-cost monitoring.

THINK   -> analyze brief, history, and unfinished questions
EXECUTE -> modify code/config, dry-run, launch training on GPU
MONITOR -> check PID, GPU, and logs with no LLM calls
REFLECT -> parse results, compare with baselines, choose next step
REPEAT  -> continue 24/7 with compact rolling memory

The architecture follows a leader-worker pattern: the leader keeps cycle-level reasoning coherent, while specialized idea, code, and writing workers stay stateless and cheap. Memory is split into a frozen project brief and a compact rolling log so the context size stays stable even after weeks of operation. During long training windows, the agent does not poll the LLM. It simply checks process liveness, GPU state, and recent log tails, which keeps runtime cost practical for real research usage.

⚙️ Quickstart

The repository is designed so a researcher can hand the setup to an AI assistant and start from a single project brief.

git clone https://github.com/Xiangyue-Zhang/auto-deep-researcher-24x7.git
cd auto-deep-researcher-24x7
pip install -r requirements.txt
python install.py
python -m core.loop --check

After installation, the user prepares a PROJECT_BRIEF.md describing the goal, codebase constraints, and experiment directions. The agent then uses that brief as the long-term task definition and keeps iterating automatically. The repository also ships with an AI guide so first-time users can paste one file into Claude, ChatGPT, or Codex and let the assistant walk through setup step by step.

🔗 Links