Neural Information Processing SystemsNeurIPS 2026
Act-onomyA Taxonomy of AI Agent Behavior

Explore Act-onomy

Browse the vocabulary, then follow the evidence in real agent trajectories.

THE VOCABULARY

Explore the Taxonomy

Start with an action. Follow it down to a concrete behavior.

Download JSON
ACT·ONOMY / AT A GLANCE
100%
Act-onomy: a map of agent behavior10 actions branch into 46 subactions and 120 leaf categories. Choose an action or subaction to explore its categories. Rotation pauses while you hover or use the keyboard.

Slowly rotating Hover to pause Drag to rotate Click to explore

10 actions → 46 subactions → 120 leaf categoriesBrowse the full codebook
120 leaf categories
Released taxonomy · v1.0 / V4.2View source
BEHAVIOR IN CONTEXT

Explore Agent Trajectories

Explore two SWE-agent tasks and a SWE-chat session shaped by human feedback.

SWE-AGENT · GPT-4 TURBO
Full report
BEHAVIOR ACROSS THE TRAJECTORYChoose a turn to inspect its evidence

AGENT THOUGHT
ACTION
View preceding observation 

      
Labels describe observed behavior; they are not a quality score. Case-study labels include extensions to the released taxonomy.Annotated JSON ↓

    

    
Look beyond the final answer. In the requests trace, the agent runs a reproduction script and inspects the result. In the django trace, it submits after acknowledging that it has not verified the change.