AgentCommander

Orchestrating AI Agents for Iterative Scientific Research.

📧 Contact: miaoxin.liu@u.nus.edu | 📖 Documentation: Read Online

---

Motivation

Born from the complex computational needs of scientific research, AgentCommander addresses a critical bottleneck in machine learning: the exhaustive cost of manual trial-and-error.

I attempted to iterate and optimize machine learning code using various existing tools, but found them lacking in flexibility. Cursor Agent excels at code completion but cannot design long-term evolutionary paths. OpenEvolve/AlphaEvolve offers powerful population-based evolution but focuses on group behavior rather than deep, customized single-agent optimization.

AgentCommander fills this gap. It is built on the belief that repetitive iteration is a task for machines, not humans. By automating the debugging and refinement cycle with a highly customizable graph-based workflow, AgentCommander empowers researchers to focus on high-level creative pursuits and systemic design.

What is AgentCommander?

AgentCommander was born from the actual demands of scientific research.

Refined through rigorous practical application, it is a graph-based workflow engine designed to orchestrate AI Agents for complex, iterative tasks. Built to leverage the diverse ecosystem of LLM CLIs (Gemini, Qwen, Claude, OpenCode, etc.), it enables Machine Learning engineers to construct highly customizable, infinite-loop workflows.

Unlike "black-box" agents, AgentCommander prioritizes Human-Centric Evolution. You define the search space and evaluation logic; the agent handles the exhaustive execution loop.

Agent Capabilities (The Loop of Discovery)

Inside the workflow, the AI acts as an autonomous researcher, capable of:

• Hypothesis & Reasoning: Analyzing current code and historical results to formulate logical improvements.
• Autonomous Coding & Debugging: Implementing changes in strategy.py and iteratively fixing errors based on execution logs.
• Multimodal Visual Feedback: "Seeing" and interpreting generated plots (e.g., loss curves, scatter plots) to detect qualitative issues like overfitting or bias.
• Meta-Learning & External Inspiration: When stuck, the agent can trigger an online search to find fresh inspiration from Arxiv papers or GitHub repositories, helping it break through local optima.
• Knowledge Evolution: Learning from both success and failure. The system extracts "Lessons" from past attempts and persists them into history.json, allowing the agent to refine its strategy and evolve across generations.

Core Features for Humans

AgentCommander provides a high-level control plane for researchers to steer the evolution:

• Hierarchical Workflows: Orchestrate macro-level evolutionary strategies (Outer Loop) and micro-level experiment execution (Inner Loop). The Progress Tree Visualization allows you to monitor the overall lineage and evolutionary status across branches and generations at a glance.
• Transparent Observability: Every experiment is isolated in its own folder. For granular details—such as generated code, multimodal outputs, or execution history—the built-in File Management page provides direct access to every artifact without leaving the UI.
• Visual Editor & High-Freedom Design: A drag-and-drop interface (assisted by AI) that offers total freedom in defining your system's logic. You can precisely control the internal lifecycle of each experiment and orchestrate complex evolutionary paths on the Progress Tree. This architecture enables advanced behaviors like cross-pollination between branches—allowing different experimental lineages to share insights and "lessons," mimicking the collaborative and non-linear nature of scientific discovery.
• Multi-Model Support: Native integration with Gemini, Qwen, Claude, and OpenCode CLIs. Use the backend that best fits your research needs.
• Safety Sandboxing: Directory-level isolation with filesystem snapshots and automated rollback.
• HPC Support: Built-in templates for PBS/Slurm clusters (e.g., NUS Vanda server).

Adapt to Your Research in Minutes

The Auto-Setup Wizard makes it easy to integrate AgentCommander into your existing workflow without rewriting your code.

1. [Case: You only have Dataset]: Scaffolds a complete project (splitting, strategy, and evaluator) from raw data.
2. [Case: You have Training Code]: Instantly adapts your existing scripts into the agent system by adding a simple interface for weight loading and evaluation.

Quick Start

1. Install Prerequisites:
   • Python 3.10+
   • LLM CLI: npm install -g @google/gemini-cli@latest (or qwen, claude, opencode-ai)
2. Clone & Install:

   git clone https://github.com/mx-Liu123/AgentCommander.git
   cd AgentCommander
   pip install -r requirements.txt

3. Launch:

   bash run_ui.sh

   Open http://localhost:8080, go to the Experiment Setup tab, and scaffold your first project.

---

Licensed under the Apache License 2.0.