Act*ONOMY: a NeurIPS 2026 taxonomy with 10 actions and 120 categories to decode agent behavior
DanielKhashabi · x · 2026-11-10
A Johns Hopkins/Caltech/Northeastern/HKU team will present ActONOMY at NeurIPS 2026 — a taxonomy for describing and analyzing AI agent runtime behavior, addressing the problem that long Claude Code/Codex-style traces are hard to interpret at scale.
- Three-level structure: 10 actions (planning, reasoning, executing, evaluating, reflecting…), 46 subactions, 120 leaf categories
- Built via Grounded Theory; every behavior label is linked to quoted evidence in the agent trace
- Open repository ships a living codebook, annotated trajectories/datasets, an automated analysis pipeline, and an extension protocol
- Enables comparing behavioral profiles across agents and characterizing one agent's behavior across runs, surfacing failure patterns
Paper, code, and dataset are all public.
More from coding & agent
- CoreWeave Launches Claude Code Skill to Build GPU Clusters and Serve Models from the CLI — altryne · 2026-11-10
- Designer ships portfolio built entirely with Claude Code, declaring the Design Engineer era — evilrabbit_ · 2026-11-10
- Stripe's claimable sandboxes: frictionless test accounts for AI agents, 50k+ provisioned daily — dl_weekly · 2026-11-10
- Simon Willison Hides AI-Generated PR Descriptions Behind Collapsed details Tags — tdhopper · 2026-11-10
- Ex-film composer vibe codes his dream ambient DAW in a week with Claude Fable 5.2 — yawnxyz · 2026-11-10
- Devs debate: Claude makes customizing enterprise software viable, exposing SaaS misaligned incentives — ivan_bezdomny · 2026-11-10