Act*ONOMY: a 120-category behavior taxonomy for annotating AI agent trajectories, at NeurIPS
DanielKhashabi · x · 2026-11-10
- A team will present Act\ONOMY at NeurIPS 2026 in Sydney: a taxonomy for understanding and analyzing AI agent behavior, developed via Grounded Theory.
- It offers a shared vocabulary of 10 actions, 46 subactions, and 120 behavior categories, linking behavior labels to exact quotes in agent traces so researchers can compare agents, inspect runtime behavior, and investigate failure patterns.
- They also released a skill for annotating long agentic trajectories into action/behavior labels, useful for both quantitative behavioral analysis and manual debugging. The accompanying figure shows a 13-turn successful SWE-agent trajectory fixing a pylint bug, with labeled phases like confirming the bug, recovering from a failed fix, and pinpointing the regex root cause.
Related event: Act*ONOMY: A Taxonomy Decomposing AI Agent Behavior into 120 Categories(2 posts)→
More from coding & agent
- 'Tell your agent to find all implicit state machines' — and neither dev knows what one is — RajaswaPatil · 2026-11-10
- Codex runs in full-color smart glasses weighing under 50 grams — Scobleizer · 2026-11-10
- Meta's Astryx design system hits 20.4k GitHub stars as agent-ready release — abhishek__AI · 2026-11-10
- Meme: what your coding subagents do when the task only needs 5 lines of code — venturetwins · 2026-11-10
- Aligning KV-cache blocks with agent turns yields up to 3x throughput gains — heghbalz · 2026-11-10
- After One Day With EnConvo, This User Uninstalled 4-5 Separate AI Tool Subscriptions — lxfater · 2026-11-10