IceShoreRadar AI × Software Maintenance

IceShore Radar AI × Software Maintenance

November 2026

[GOOD NEWS]

15.8%

less agent execution time

Catching coding agents' invalid commands before they run cut task time by 15.8%

A preprint describes StateMirror, which checks an agent's planned commands against the live repository before they run, catching references to missing files or packages. In 4,398 logged sessions such errors cost a median of three extra steps. On 79 live tasks it cut execution time 15.8% and token use 11.8% without lowering the solve rate.

・ Chinese public university · preprint

Source: Sun Yat-sen University (arxiv.org)

Photo: Markus Spiske markusspiske / Wikimedia Commons, CC0

[UPDATE]

15.2%

backed up before changes

AI coding agents back up files before risky changes only 15.2% of the time

Researchers built 205 tasks from cases where users said they regretted not saving something an agent later changed. Six frontier models finished 86.1% of tasks on average but backed up the affected files, environments or settings beforehand only 15.2% of the time. An add-on backup layer the team built covered every such action.

・ US private university · preprint

Source: Northeastern University (arxiv.org)

Photo: ajay_suresh / Wikimedia Commons, CC BY 4.0

[UPDATE]

59.6%

top projects with AI instruction files

AI agent instruction files now appear in 59.6% of the most-starred GitHub projects

A preprint study of 22,854 widely used, active GitHub projects found instruction files for coding agents, AGENTS.md or CLAUDE.md, in 26.3% of them and in 59.6% of the most-starred. Of 341,852 statements analyzed, 74.0% of directives gave no reason, and projects split on whether to disclose AI use.

・ Researcher with no listed affiliation · preprint

Source: Pawel Zarzycki (arxiv.org)

Photo: Marc Mueller seven11nash / Wikimedia Commons, CC0

[GOOD NEWS]

66.1%

cost cut vs. earlier method

New AI method ports whole C projects to Rust at 66.1% lower cost than prior work

SpeC2Rust, described in an arXiv preprint, has an AI model write specifications for each module of a C codebase, then translate and repair the code against them. All eight test projects compiled in Rust and passed their original end-to-end tests with no unsafe code. Against an earlier AI-based approach, cost fell 66.1% and input tokens 84.7%.

・ Chinese public university · preprint

Source: Harbin Institute of Technology, Shenzhen (arxiv.org)

Photo: Fumikas Sagisavas / Wikimedia Commons, CC0

[UPDATE]

13,919

translation errors labeled

Preprint catalogs 13,919 mistakes AI models make when translating code between languages

Researchers had seven AI models translate code across six language pairs on three benchmarks and hand-labeled 13,919 cases where the output strayed from the original. They sort the errors into three types, including semantic conflicts and lost project context, and propose a fix that retrieves similar past translations.

・ Chinese university · preprint

Source: Nanjing University (arxiv.org)

Photo: Kevin Dooley / Wikimedia Commons, CC BY 2.0

[UPDATE]

AI as a tool

Torvalds on AI in Linux work

Linus Torvalds says Linux wants competent developers to use AI as a tool

Linus Torvalds, creator of the Linux kernel, told a Register columnist that arguments against using AI for commit messages are misguided, and that Linux wants capable developers to use AI as a tool. He said submitters who send AI-made patches they don't understand should be discouraged. The Register published the remarks on Oct. 23.

・ UK tech news site · column

Source: The Register (theregister.com)

Photo: Elizabeth Krumbach Joseph / Wikimedia Commons, CC BY 2.0

[BLIND SPOT]

15.4%

failures fixed in 2 tries

When requirements shift, coding AIs recover from just 15.4% of failures in two tries

A preprint introducing the FeatureLoop benchmark tested 12 frontier models on 61 human-reviewed tasks from 34 repositories, with requirements changing mid-project. The best model finished 72.1% of tasks, but only 15.4% of failed stages recovered within two rounds of corrective feedback. The authors also saw incomplete repairs and regressions.

・ Major Chinese tech company · preprint

Source: Tencent (arxiv.org)

Photo: 花浅葱 / Wikimedia Commons, CC0

[UPDATE]

14.0%

specs fully delivered

Coding agents fully delivered at most 14% of enterprise product specs in a new test

A benchmark of 43 tasks in mature enterprise codebases, checked against 750 expert-audited points, measures whether coding agents fulfill full product requirement documents. Across 11 model and tool setups, the best complete-delivery rate was 14.0%, by OpenAI's GPT-5.6 Sol; missing integration was a main cause.

・ Chinese public university · preprint

Source: Zhejiang University (arxiv.org)

Photo: 江上清风1961 / Wikimedia Commons, CC BY 3.0

[UPDATE]

Teams ×AI bug fixes

bug report to fix

GitHub's Copilot can now fix bugs reported in Microsoft Teams and propose the change

GitHub said on Oct. 23 that its app for Microsoft Teams now brings in Copilot, its AI coding assistant. When a bug report lands in a Teams chat, users can mention GitHub to have Copilot investigate, fix the bug and open a pull request for the team to review. The old GitHub app for Teams will be retired on Nov. 16.

・ US developer platform · changelog

Source: GitHub (github.blog)

Photo: Simon Hattinga Verschure webmarbles / Wikimedia Commons, CC0

[WARNING]

AI-written code ×global outage

Forrester's forecast for 2027

Forrester predicts a bug in AI-written code will cause a global outage in 2027

Research firm Forrester released its 2027 technology and security predictions on Oct. 21. It expects AI-assisted coding and patching to outpace teams' capacity to test, and forecasts that a bug or regression in AI-generated code will trigger an outage spanning several industries worldwide.

・ US research firm · press release

Source: Forrester (forrester.com)

Photo: Carl Lender from Sunrise, USA / Wikimedia Commons, CC BY 2.0

[UPDATE]

61 defects

defects found in migration

Checks for AI-driven COBOL-to-Java migration catch 61 defects but miss rounding errors

Aimed at agent-driven migrations, a preprint generated 849 end-to-end test cases for a 33,429-line COBOL program to check its Java translation, finding and fixing 61 defects in the code generator. When the authors planted 450 defects, the check caught 47.6%, including every deleted write in executed code, but no rounding errors.

・ US financial services company · preprint

Source: American Express (arxiv.org)

Photo: giggel / Wikimedia Commons, CC BY 3.0

[UPDATE]

AI debugging ×code comes last

gather evidence before code

Martin Fowler backs letting AI debugging agents read the code last, not first

On his site Oct. 21, software author Martin Fowler recommended an essay by Cindy Sridharan on debugging with AI agents. Her view, which he summarized approvingly, is that an agent should first gather evidence about how the system behaves and only then form hypotheses from the code. Fowler noted that a human is still needed to steer the agent.

・ British software developer · personal site

Source: Martin Fowler (martinfowler.com)

Photo: Paris Buttfield-Addison / Wikimedia Commons, CC BY 2.0