ACT*ONOMY: A Taxonomy for Analyzing Autonomous Agent Behavior
A new taxonomy named ACT*ONOMY has been unveiled by researchers to categorize and evaluate the runtime behavior of autonomous agents such as Claude Code and Codex, which can now function for extended periods, ranging from hours to days. This taxonomy, created using Grounded Theory, organizes agent activities into a three-tier hierarchy consisting of 10 actions, 46 subactions, and 120 leaf categories. Its purpose is to assist humans in understanding complex natural-language reasoning paths and execution traces on a large scale, facilitating the identification of inefficiencies, bug resolution, and enhanced oversight. The initiative features an open repository containing a dynamic taxonomy and an automated analysis pipeline for agent trajectory application. Detailed findings are available in arXiv:2605.13625.
Key facts
- Autonomous agents such as Claude Code and Codex now operate for hours or days.
- ACT*ONOMY combines Action and Taxonomy.
- The taxonomy has a three-level hierarchy: 10 actions, 46 subactions, 120 leaf categories.
- Developed through Grounded Theory.
- Includes an open repository with a living taxonomy and automated analysis pipeline.
- Aims to help interpret unstructured natural-language agent trajectories.
- Supports downstream tasks like diagnosing inefficiencies, fixing bugs, and oversight.
- Published on arXiv with ID 2605.13625.
Entities
Institutions
- arXiv