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

What is Act-onomy?

Act-onomy is a taxonomy for describing and analyzing AI agent behavior at runtime. It provides a shared vocabulary of 10 actions, 46 subactions, and 120 leaf categories to help people make sense of what agents do.

Read the trace

“To verify that this change resolves the issue, we should create a script that replicates the problem as described.”

SWE-agent · Turn 6

Describe the behavior

ActionPlanningSubactionDecompose taskLeaf category

Decompose into subgoals with success conditions

Follow the trajectory

06

Plan a checkPlanning

08

Run the scriptExecuting

09

Check the resultEvaluating

A real agent trace, described at three levels and examined in context.

Agents produce long reasoning trajectories and execution traces, but these records are often difficult to interpret at scale. Act-onomy connects behavior labels to quoted evidence, helping researchers, agent designers, and users examine how an agent plans, reasons, executes, evaluates, and adapts.

The taxonomy was developed through Grounded Theory. The accompanying repository contains the living codebook, research datasets, automated analysis tools, and a protocol for extending the vocabulary as new agent behaviors emerge.

Act-onomy paper (Gao et al., NeurIPS 2026)Code, data, and analysis tools

Overview

  • Purpose — Describe and analyze agent behavior at runtime.
  • Structure — Action → Subaction → Leaf category.
  • Evidence — Each trace annotation is linked to the words behind it.
  • Applications — Compare behavioral profiles across agents and examine differences within a single agent’s trajectories.
  • Extension — Adapt the codebook to behaviors in new agent systems.

The released taxonomy

  • 10 actions, including Planning, Reasoning, Executing, and Reflecting.
  • 46 subactions that distinguish more specific kinds of behavior.
  • 120 leaf categories with definitions or examples from the codebook.

Getting Started

An example from a real SWE-agent trajectory. Switch between the raw text and its annotation, or follow the next steps:

SWE-agent / requests-2317REAL TRACE
TURN 06 · VERIFY

The same trace. A new way to read it.Explore

Browse the complete taxonomy or explore three annotated trajectories: two SWE-agent tasks and a SWE-chat session with human feedback (best viewed on desktop).

Browse the Act-onomy taxonomyExplore annotated agent trajectories

Download the released codebook and machine-readable taxonomy:

Download the codebook (CSV)Download the taxonomy (JSON)

How Was It Developed?

Act-onomy uses a Grounded Theory approach to turn descriptions of agent behavior into a structured, extensible vocabulary.

  1. Ground the vocabulary. Collect behavior descriptions from agent research, review their evidence, and refine category boundaries.
  2. Apply it to trajectories. Link labels to quoted evidence and examine patterns across agents and across runs.
  3. Keep it open to change. Propose new behaviors through a documented extension protocol.
Read about the methodology

Tools and Resources

Use the released analysis skill to annotate an agent trajectory and generate an evidence-linked report:

Analyze your own trajectories

Explore the research corpus, examine behaviors in new papers, or extend the codebook:

Explore the research datasetsDiscover behaviors in new papersRead the extension protocol

Cite This Work

How to Interpret Agent Behavior
NeurIPS 2026

Jie Gao, Kaiser Sun, Jen-tse Huang, Katherine Van Koevering, Sijie Ji, Heyuan Huang, Weiyan Shi, Zhuoran Lu, Ziang Xiao, Daniel Khashabi, and Mark Dredze

View citation


Have Questions?
Ask a question or share feedback through the Act-onomy GitHub repository.