SuperAPI AI Directory
Partner with us中文AI newsFavorites
All toolsAI searchauto-deep-researcher-24x7

auto-deep-researcher-24x7

Auto Deep Researcher 24x7 is an autonomous AI agent for deep learning experiments that can automatically run experiments 24/7 while you sleep. It uses a Leader-Worker architecture with a constant-size two-layer memory, and significantly reduces overhead through zero-cost monitoring (no LLM calls during training), with the official claim of about $0.08 per 24-hour cycle. The work loop is THINK → EXECUTE → REFLECT: read the project brief, analyze historical results, plan the next experiment, modify code/config and do a dry run before launching training on the GPU, monitor the training process at zero LLM cost, parse results, compare with the baseline and decide the next step, then repeat. It supports local/SSH/Slurm execution backends, and supports access to Chinese LLM APIs such as DeepSeek, Qwen (dashscope), Kimi (moonshot), and GLM (zhipu) through provider presets. It has a built-in experiment ledger, stagnation signals, research logs, violation scanning and phase gates, anti-burn rate limiting, code understanding and literature retrieval tools, and supports exporting progress to Obsidian or local text.

Open website

About this tool

Purpose

Auto Deep Researcher 24x7 is an autonomous AI agent for deep learning experiments, with the goal of automatically performing repetitive experiment operations work so researchers can invest their time in thinking and judgment. It is suitable for people who already know what experiment they want to run but do not want to watch the training loop the whole time: launch training, decide the next variant, and keep iterating while you sleep. The official description explicitly states that it does not replace the researcher, but takes over the "experiment operations layer."

Main capabilities

The agent works according to the THINK → EXECUTE → REFLECT loop:

  • Think: read the project brief, analyze previous results, and plan the next experiment.
  • Execute: modify code/config, perform a dry-run, and then launch training on the GPU.
  • Monitor: monitor training at zero LLM cost (only process checks and log reading).
  • Reflect: parse results, compare with the baseline, and decide what to try next.
  • Repeat: run 24/7 without human intervention.

it uses a Leader-Worker agent system, with only one worker running at a time and the rest idle at zero cost; the two-layer memory remains constant in size. The official description emphasizes that monitoring uses a zero-cost strategy: during training the agent sleeps and makes no API calls, and only wakes the LLM when training ends; for a 24-hour cycle with 8 hours of training, the LLM call cost is about $0.08.

How to use

The official shortest path is three steps:

  1. Create a project folder containing a file PROJECT_BRIEF.md.
  2. Run /auto-experiment --project /path/to/project --gpu 0.
  3. Use /experiment-status to check progress, or use the optional Obsidian/local text record.

You can also open AI_GUIDE.md and let Claude / ChatGPT / Codex guide you through setup. The installation steps have API key requirements, for example obtaining and setting it at console.anthropic.com; the official description also mentions direct Python startup and the Claude Code method.

You can control the research direction through three files: PROJECT_BRIEF.md (stable goals, constraints, allowed search space), HUMAN_DIRECTIVE.md (temporary redirection for the next cycle), workspace/MEMORY_LOG.md (rolling memory of results and decisions). Ways to check progress include /experiment-status、/progress-report、/obsidian-sync, as well as local text records saved to workspace/progress_tracking/ when no Obsidian vault is configured.

Recent Version Highlights (according to official update notes)

  • 2026-06-03: Added domestic LLM API presets, setting agent.provider to the one-word presets deepseek, qwen (dashscope), kimi (moonshot), or glm (zhipu), automatically filling in the OpenAI-compatible base_url and default key environment variables (DEEPSEEK_API_KEY / DASHSCOPE_API_KEY / MOONSHOT_API_KEY / ZHIPUAI_API_KEY), requiring only setting that vendor's model id; base_url / api_key_env can still be overridden for self-hosting or proxy endpoints.
  • 2026-06-02: Added the Slurm execution backend (execution.mode: "slurm"), with the controller remaining local and training submitted to the login node via a one-off SSH call to sbatch --parsable; sacct serves as the sole authority for liveness status, and two boundaries guarantee termination of the monitoring loop. Also added "real experiment outcomes": monitoring obtains the true final state of completed jobs via final_status(), and FAILED / TIMEOUT / CANCELLED are no longer silently recorded as completed; results flow into state.json, the experiment ledger, and the REFLECT context; pid-only backends (local / ssh) are reported as indeterminate and keep their original behavior.
  • 2026-06-01 v2.0: Added the autonomy layer (experiment ledger experiments.jsonl, data-driven stagnation signals, append-only research logs DEAD_ENDS.md and INSIGHTS.md, zero-cost violation scanning and advisory stage gates, proactive anti-burn-rate limiting agent.max_cycles_per_hour); added agent tools (search_code, list_tree, read_file line ranges, get_paper, search_arxiv), and noted that all new tools behave consistently under local and SSH execution modes.

Applicable Scenarios and Limitations

Applicable scenarios: deep learning research that already has clear experimental goals and wants automated repetitive experiment operations (launching, monitoring, parsing logs, hyperparameter sweeps). The official materials also express usage expectations: do not use this tool for fabricating results, for "research" without a human in the loop, or to bypass the parts that require genuine human understanding; science should remain pure, and ideas, explanations, and responsibility should be left to humans.

Limitations and caveats: the official materials do not state specific benchmark results (the original text says "Battle-Tested Results" come from months of 7×24 real-world operation rather than benchmarks); they do not list supported models, GPU model requirements, system dependencies, license agreements, or pricing policies. The cost figure (about $0.08 per 24-hour cycle) is the official statement under their setup (8 hours of training) and may vary in practice depending on configuration. Installation and operation require an API key; the specific way to obtain and set it is subject to the official documentation.

Related tools

Browse category