Preprint argues code is the operating infrastructure for AI agents

The survey argues that in agentic systems, code has shifted from being a target artifact to an "operational substrate" for reasoning, acting and verification. It offers a unified view where code constitutes the basis for agent infrastructure, termed "code as agent harness.

By Nadia Vale

November 7, 2026 at 9:04 PM EST1 min read

Key takeaways (3)

1.The arXiv abstract frames code as the basis for agent infrastructure and organizes the survey into interface, mechanisms and multi-agent layers.
2.Stanford, Meta and UIUC authorship comes only from an X post; the retrieved arXiv text showed no authors or affiliations.
3.The abstract lists open challenges including evaluation beyond task success, shared state across agents and human oversight of safety-critical actions.

Key takeaways (3)

1.The arXiv abstract frames code as the basis for agent infrastructure and organizes the survey into interface, mechanisms and multi-agent layers.
2.Stanford, Meta and UIUC authorship comes only from an X post; the retrieved arXiv text showed no authors or affiliations.
3.The abstract lists open challenges including evaluation beyond task success, shared state across agents and human oversight of safety-critical actions.

A preprint posted to arXiv, "Code as Agent Harness," argues that code now serves as the operating infrastructure for AI agents rather than only their output.

The retrieved arXiv text did not show authors, affiliations or a submission date. An X post by @Gyome1_2, published Nov. 7, 2026, attributes the survey to researchers at Stanford, Meta and UIUC. The post is a secondary summary and does not establish that attribution.

According to the arXiv abstract 1, recent large language models are strong at understanding and generating code, from competitive programming to repository-level software engineering. In emerging agentic systems, the authors write, code "increasingly serves as an operational substrate" for reasoning, acting, environment modeling and execution-based verification. They call this view "code as agent harness."

The survey is organized around three layers. The first is the harness interface, where code connects agents to reasoning, action and environment modeling. The second is harness mechanisms: planning, memory and tool use for long-horizon execution, plus feedback-driven control and optimization. The third is scaling from single-agent to multi-agent settings, where shared code artifacts support coordination, review and verification.

The abstract lists applications spanning coding assistants, GUI/OS automation, embodied agents, scientific discovery, personalization and recommendation, DevOps, and enterprise workflows. Its open challenges are evaluation beyond final task success, verification under incomplete feedback, regression-free harness improvement, consistent shared state across agents, human oversight for safety-critical actions, and multimodal extensions.

The X post summarizes a harness as "the tools, memory and checks around the model" and gives the formula "agent = model + harness." The abstract does not contain that formula or definition. The post's further list of requirements (permissions, a sandbox, hooks and replayable logs) also does not appear in the abstract.

Automatica has separately covered an X post describing AgentDojo 3, a testing environment for agent security and reliability. That coverage does not concern this survey.

The paper is a preprint. This report is based on the abstract, because the PDF and HTML versions could not be retrieved. The abstract's comments field points to a GitHub repository, but the retrieved text did not include the link.

Sources3 documents

1. arxiv.org

arxiv.org/abs/2605.18747

6 checked claims

2. x.com

x.com/Gyome1_/status/2119027969884586336

3 checked claims

3. automaticapress.com

automaticapress.com/articles/unconfirmed-x-post-describes-ai-security-project-agentdojo-no-primary-source-pro-mw6132yo

1 checked claim

Citations link to the documents this article draws on; they do not imply independent corroboration.

Preprint proposes unified framework for judging mechanistic interpretability claims

An arXiv preprint evaluates whether mechanistic interpretability techniques actually make AI systems safer by surveying methods for analyzing internal model computations. The paper proposes a unified evaluation framework and releases code notebooks for reproducibility.

Nadia Vale · about 9 hours ago

METR estimates Anthropic's benchmark progress is about twice as fast

METR’s assessment points to a material rise in Anthropic’s research speed and says AI assistance is likely the main driver. But Anthropic’s redactions mean the most informative supporting material is not public, limiting independent evaluation of the claim.

Julian Park · about 14 hours ago

About 40% of AI-using scientists expect AI to complete full research projects by 2030

Anthropic surveyed AI-using scientists and analyzed session data, finding that 40% expect autonomous research by 2030, but current usage concentrates on data analysis rather than idea generation, which scientists consider most important.

Adrian Cole · about 15 hours ago

Sources3 documents

More from Automatica Press

Preprint proposes unified framework for judging mechanistic interpretability claims

METR estimates Anthropic's benchmark progress is about twice as fast

About 40% of AI-using scientists expect AI to complete full research projects by 2030