Agent Capability

AgentCommander empowers AI agents to mimic the cognitive loop of a human researcher: Hypothesize, Observe, Reference, and Summarize.

By orchestrating these cognitive steps into a structured workflow, we achieve a highly efficient, automated trial-and-error engine.

The Default Research Cycle

In the default configuration (default_graph.json), every time a new experiment folder is created, the Agent follows a rigorous scientific process:

1. Hypothesis Generation

• Context: The agent reads the current code, previous execution logs, and critically, performs Visual Analysis on any result plots (e.g., loss curves, prediction scatter plots).
• Action: Based on this multimodal input, it formulates a specific hypothesis for the current experiment (e.g., "The model is overfitting, so I will increase dropout").

2. Implementation & Debugging

• Coding: The agent modifies the code (strategy.py) to test the hypothesis.
• Observability: It proactively adds debug print statements to gather more information during execution, just like a human developer would.
• Execution: The code is run against the immutable evaluator.py.

3. Summary & History

• Record: Regardless of success or failure, the agent summarizes the experiment design and performance into history.json.
• Evolution: If the score improves, it advances to the next level. If it fails, it retries until a limit is reached.

4. Meta-Reflection (Getting Unstuck)

• Memory: When starting a new experiment, the agent reads the history of previous attempts.
• External Inspiration: If the system detects a stagnation (multiple generations without breakthrough), it triggers a special Meta-Analysis step. The agent uses Online Search tools to find relevant papers (Arxiv) or open-source repositories (GitHub) to find fresh inspiration.