SINGAPORE, 8 MAY 2026 — Global git push volume — the technical action by which developers upload completed code changes to shared repositories — rose 78 per cent year on year, according to Microsoft, in one of the clearest quantitative signals yet that AI coding tools have crossed from novelty to infrastructure.
Disclosure: RECATOOLS is written with the assistance of Claude, made by Anthropic, which is also named in this article. Readers should weigh that interest.
The figure, published by Microsoft in its State of Global AI Diffusion report on 7 May 2026, measures global push volume rather than AI usage or the output of individual developers. Git pushes are a proxy for code being shipped: code that has been written and committed to a version control system for integration. Microsoft said the increase was reflected in production by Anthropic's Claude Code, OpenAI's Codex and Microsoft's GitHub Copilot, and noted that, for now, AI coding capabilities "may be increasing demand for the employment of software developers" — US software developer employment reached a record of about 2.2 million in 2025.
The Tools Driving the Surge
Claude Code, Anthropic's terminal-native coding agent, is designed for complex, multi-file software development tasks. Unlike IDE-based completions that suggest the next line or function, Claude Code operates on entire codebases — planning feature implementations across multiple files, writing tests, debugging failures, and committing changes to Git with descriptive commit messages. It runs from the terminal in a project directory, without requiring an IDE plugin.
GitHub Copilot, now in its Business and Enterprise tiers, has achieved broad corporate deployment as the institutionally approved AI coding tool for large organisations. Its integration into Visual Studio Code and the JetBrains IDE family — the two most widely used professional development environments — means that AI code suggestions are now present by default in the development environments of millions of enterprise software engineers. GitHub Copilot Enterprise's ability to index an organisation's private repository history and suggest code consistent with internal conventions and libraries is specifically valuable for large codebases where consistency matters as much as productivity.
OpenAI's Codex, the third tool Microsoft named, adds to the same trend. Together, these tools form a layered AI coding stack — agents for multi-step tasks alongside inline completion in the editor — that is now common in professional software development.
From Wrappers to Autonomous Systems
The nature of what developers are building with AI tools has shifted significantly over the past twelve months. The 2024 wave of AI application development was characterised by "wrappers" — products that placed a user interface over an LLM API and delivered value primarily through the model's capability rather than the product's own logic. Those products are now under severe competitive pressure as the underlying models improve and the differentiation of a thin wrapper over a common API diminishes.
The 2026 wave is characterised by autonomous systems — applications that use AI agents to manage extended workflows across multiple platforms without requiring constant human prompting. Trending GitHub repositories in April and May 2026 include "ruflo," an orchestration layer for Claude, and "open-design," a local-first alternative to Anthropic's design tools. "OpenClaw" agents (version 2026.5.2, released 3 May 2026) are designed for continuous operation across WhatsApp, Slack, and Discord, maintaining context and taking actions across platforms without human input at each step — what the developer community has begun calling "local AI sovereignty."
This architectural shift from wrapper to autonomous system is reflected in the git push statistics. A developer building an autonomous system is producing significantly more code than a developer building a wrapper: the orchestration layer, the tool integrations, the state management, the error handling, and the testing infrastructure for an autonomous agent collectively represent substantially more code than a simple API integration. The 78 per cent surge in git pushes is partly a story about individual developer productivity gains, but it is also a story about the increasing complexity of what developers are choosing to build.
Quality and the Hidden Risk
The 78 per cent increase in code production raises a question that the productivity metrics alone do not answer: is quality keeping pace with quantity? The research literature on AI-assisted code quality is nuanced. Controlled studies have found that AI-generated code tends to be syntactically correct and functionally adequate for well-specified tasks, but is more likely to introduce subtle logical errors, inadequate error handling, and security vulnerabilities than equivalent code written by experienced developers working without AI assistance.
The security vulnerability concern is quantitatively significant. Stanford research first published in 2022 found that participants with access to an AI coding assistant wrote significantly less secure code than those without one — and were more likely to believe that their code was secure. When git pushes increase by 78 per cent, and a meaningful proportion of that code is AI-generated, the aggregate security vulnerability surface of deployed software is expanding at a rate that security review processes built for human-pace development cannot match.
CVE-2026-3854, a high-severity remote code execution vulnerability in GitHub Enterprise Server (CVSS 8.7 under v4.0) that was patched in March and detailed publicly in late April, illustrates that development infrastructure is itself a target. As AI coding tools drive more git push volume through the same server-side processing pipelines, the attack surface of the development infrastructure itself becomes a higher-value target. The productivity gain and the security exposure are not independent variables — they scale together.
Singapore: GovTech, Startups, and the Talent Shift
The AI coding productivity surge has direct implications for how Singapore's technology employers think about team sizing, hiring, and skills requirements.
A 2024 study of GitHub Copilot by GovTech Singapore's Engineering Productivity Programme reported that coding and task speed increased by 21 to 28 per cent. Findings of that kind are likely to inform how Singapore government technology functions approach AI coding tool adoption over the next 12 to 18 months.
For Singapore's private sector technology companies, the productivity dynamics create a structural shift in hiring economics. If AI coding tools let each developer ship more, a company's engineering output can scale without commensurate headcount growth — although the 78 per cent rise in global push volume is an aggregate figure, not a measure of what any individual developer produces.
In our view, training programmes such as IMDA's TechSkills Accelerator will need to reflect the AI coding reality. The skills that differentiate a productive AI-era developer from an unproductive one are not the ability to write boilerplate code quickly — AI has commoditised that — but the ability to specify requirements precisely, review AI-generated code critically for security and logical correctness, architect systems that AI agents can implement reliably, and debug the failures that autonomous coding agents produce at the boundaries of their competence.
DevSecOps at the Scale of AI Output
The security imperative created by the 78 per cent code production surge requires DevSecOps practices to scale at the same pace as code output. This is not trivially achievable: security review processes that were designed for human-pace code production — manual code review, quarterly penetration testing, annual security audits — cannot absorb a 78 per cent increase in code volume without either degrading review quality or expanding security team headcount commensurately.
The practical response is to automate the security review layer to match the pace of AI-assisted code production. Static application security testing (SAST) tools that scan code for vulnerability patterns as part of the CI/CD pipeline — integrated with GitHub Actions or equivalent tooling — can maintain security review throughput without proportional headcount increases. The combination of AI coding tools for production and AI-powered security scanning for review creates a virtuous cycle in which the productivity gain is not undermined by a security deficit. For developers looking to verify cryptographic outputs or hash functions in their AI-generated code, a hash generator provides a quick validation layer.
The 78 per cent surge in global git push volume is a measurement of the present. The more consequential question is what the equivalent metric will be twelve months from now, as the next generation of AI coding agents — those capable of managing entire feature development cycles autonomously — enters production use across the global developer community. The developers and organisations that build security practices capable of scaling with that trajectory will be measurably better positioned than those that treat security as a constraint on velocity rather than a prerequisite for it.
Correction, 18 October 2026: An earlier version said developers at Grab and Sea Group had described specific productivity gains from Claude Code; we have no record of those statements and they have been removed. It also overstated GovTech's use of GitHub Copilot (the published record is a 2024 study reporting 21 to 28 per cent faster coding), described the 78 per cent rise in git pushes as a per-developer output measure, gave the wrong year for the Stanford study, described the TechSkills Accelerator as an IBF programme and misstated the severity of CVE-2026-3854; these have been corrected, and unsourced claims about Singapore's tech workforce and companies' headcount have been removed.
Sources
- Microsoft — State of Global AI Diffusion Report, 7 May 2026
- Devflokers.com — AI News May 2026
- GovTech Singapore — Harnessing the Potential of Gen-AI Coding Assistants in Public Sector Software Development (arXiv 2409.17434), 2024
- Perry et al., Stanford University — Do Users Write More Insecure Code with AI Assistants? (arXiv 2211.03622), 2022
View author profile → · Editorial policy
