What this resource does
Core uses
About this resource
Deep Researcher Agent installs matching Claude Code commands and Codex Skills for iterative GPU experiments, progress tracking, paper discovery, paper analysis, conference search, and optional Obsidian notes.
This page groups representative academic components by task; review the repository for the complete inventory.
Iterative experiment operations
Edits research code, launches GPU training, monitors results, records each cycle, and proposes the next bounded variation.
- Typical inputs
- PROJECT_BRIEF.md with goal, codebase, search space, and constraints; GPU, project code, data, and stopping rules
- Output
- Experiment code, logs, ledger entries, metrics, and progress state.
Best for: Researchers who already know the experiment they want to run and need help with repetitive operations.
Components for this task
experiment-status
Reports the current experiment goal, best result, cycle count, and recent decisions.
Suggested workflow requests (2)
Run the auto-experiment workflow for the project in /path/to/project using GPU 0.
Show the current experiment status, including the best result, cycle count, and recent decisions.
Adapted into complete requests from official trigger wording in README.md.
Paper and conference tracking
Searches recent papers, analyses selected papers, and tracks relevant conference information.
- Typical inputs
- Research topic, paper identifier, or conference criteria
- Output
- Paper records, analysis notes, or conference-search results.
Best for: Keeping the experiment loop connected to current literature and deadlines.
Components for this task
Progress reporting and notes
Produces status summaries and writes optional Obsidian or local-text research notes.
- Typical inputs
- Experiment workspace and optional Obsidian vault
- Output
- A structured progress report and persistent research notes.
Best for: Reviewing current state without reading every training log.
Components for this task
Use boundaries
Limits and checks
Autonomous compute use
It can consume compute, API budget, and researcher time while following an unproductive direction.
Set cycle and hourly limits, monitor the ledger, and use human directives and stop rules.
Remote execution
A configuration or command error can affect remote research infrastructure.
Use a restricted account, isolated project directory, and reviewed command permissions.
Result validity
It may optimise the wrong metric, leak data, or overfit repeated trials.
Pre-register evaluation rules and independently rerun and inspect key results.
More boundaries
- Do not start the experiment loop without a concrete PROJECT_BRIEF.md, accessible code and data, and explicit stopping rules.
- The controller does not establish that an experimental design, metric, or claimed improvement is scientifically valid.
- One successful run or one agent-selected variation is not sufficient evidence for a research conclusion.
Technical details
- Resource type
- Toolkit
- Author or maintainer
- Xiangyue-Zhang
- Source last updated
- 3 Jun 2026
- Last verified
- 15 Jul 2026
- Documented applications
- Claude Code, Codex
- Documented AI models
- Claude, OpenAI-compatible models, DeepSeek, Qwen, Kimi, GLM
- Licence
- Apache-2.0
- Access
- Publicly available
- Additional costs
- Platform terms or usage limits may apply. API usage fees may apply for selected components. External services, software, compute, or data access may have separate costs.
- Skill instruction language
- English
- Documentation language
- English
- Repository languages
- Python, Markdown
- Dependencies
- Python 3.10 or later; Claude Code or Codex; One or more NVIDIA GPUs; Configured API or logged-in CLI provider; Git and component-specific training packages
- Review status
- Not tested
Continue exploring