Fri 13 Nov
LiveGet it free
ZOTPAPER

News without the noise

← All concepts

code as action

Have a language model write a concrete program instead of blurting out a result, then let a normal interpreter run it. Arithmetic, layout, motion, and tool calls happen exactly as computed. Vague instructions become precise, editable, reusable code. The model does the high-level reasoning; the code does the low-level action.

The papers find this separation consistently helps. Agents that generate executable scripts beat direct answering, with large accuracy gains in medical calculations and better visual composition. Code is often not written once: agents run, inspect, revise, or choose among reusable tools after each step. Formal verification also becomes possible, and frameworks expose code as a shared interface, so both people and models control behavior through familiar Python instead of task-specific abstractions.

Papers
29
Released code
2
First seen
May 2026
Latest
Nov 2026

26 papers in the last two months, against 2 in the two before.

Start here

Three papers to read first, in the order they came out
  1. 1Open agent stack improves computer tasks beyond GUI-only control

    OpenGCUA with code execution and model-driven decisions improves task scores by 19.6 points on OSWorld-v2.

    Oct 2026
  2. 2Game engine expertise and verification feedback boost AI 3D scene completion from 18% to 48%

    Language instructions can drive controllable Unreal Engine worlds through agent-generated code with verification and localized repair, preserving persistent scene and timeline state.

    Oct 2026
  3. 3Video generators can better follow physics with code-built scene drafts

    Video-Coder improves video generation by having an AI agent write and execute code to build a 3D scene before rendering the final video.

    Nov 2026

How fast it is moving

JunJulAugSeptOctNov so far

Often studied with

agent skillsLearned capabilities that enable AI agents to perform specialized tasks with improved effectiveness.7 togetheragentic evaluationMeasuring AI agents' performance in complex tasks, highlighting failures in planning, monitoring, and real7 togetheragent harnessA system that structures, verifies, monitors, or controls how an AI agent interacts with external tools, environments, or execution runs.6 together

Who is working on it

Alibaba Group 3NVIDIA 2Carnegie Mellon University 2National University of Singapore 2Microsoft Research 2Shanghai Jiao Tong University 2

The papers

Most central to this idea first, not most recent.

Concepts are extracted from each paper and reused across the corpus, so this page grows on its own as the desk reads.

code as action · Research concepts | Zotpaper