AI Now Writes 75% of Google's New Code — The Agentic Coding Era Has Arrived

The floor of software engineering is shifting. In the span of eighteen months, AI has gone from an autocomplete assistant to the default engine behind most of the new code shipped by the world's largest tech companies. At Google Cloud Next 2026, CEO Sundar Pichai revealed that 75% of all new code at Google is now generated by AI and approved by human engineers — up from 50% six months earlier, and just 25% in October 2024 . Engineers, Pichai said, are now operating in "truly agentic workflows," orchestrating autonomous digital task forces rather than writing code manually .

Google's numbers are the most dramatic public disclosure so far, but they reflect an industry-wide transformation. AI coding agents — tools that can read entire repositories, plan multi-step tasks, write code, run tests, debug failures, and propose pull requests — moved from experimental research previews to billion-dollar products in 2025 and 2026. The four major players — Google, Microsoft, OpenAI, and Anthropic — are now locked in an arms race to build the most capable, cost-efficient, and deeply integrated coding agents, and the consequences for how software is built, and for the people who build it, are just beginning to materialize.

The Four-Way Race in Agentic Coding

Google: Gemini Powers an Internal Revolution

Google's 75% figure, announced by CEO Sundar Pichai at Google Cloud Next 2026 in Las Vegas, is the most striking public benchmark for how quickly AI-generated code has scaled inside a major engineering organization . The company's internal AI coding models, Gemini, have been integrated into performance review goals — as of Q4 2025, some organizations set targets that 55% of code changes should be "Agent-Assisted," and for the first half of 2026, 65% of engineers in Google's creation org are expected to write more than 75% of their committed code using AI . The company also reported that AI-assisted code migrations were completed six times faster than they had been a year earlier with engineers working alone .

Microsoft: MAI-Code-1-Flash — An In-House Challenger for GitHub Copilot

On June 2, 2026, at Microsoft Build, Microsoft introduced MAI-Code-1-Flash . It is a 5-billion-active-parameter coding model built end-to-end by Microsoft on licensed data, with no distillation from OpenAI, Anthropic, or any other third-party model . The model uses a sparse Mixture-of-Experts architecture with 137 billion total parameters and a 256,000-token context window, and was trained from March through May 2026 inside GitHub Copilot's production harness — meaning the model learned on the same real-world developer workflows it would later serve .

The result is a model that scores 85.8% on Microsoft's adversarial coding benchmark and approximately 51% on SWE-Bench Pro, outperforming Anthropic's Claude Haiku 4.5 by 16 percentage points on the latter while consuming up to 60% fewer tokens on complex coding tasks . MAI-Code-1-Flash began rolling out to GitHub Copilot users in Visual Studio Code on June 2, across Free, Pro, Pro+, and Max plans, with third-party access available through Fireworks AI, Baseten, and OpenRouter .

OpenAI: Codex Becomes a Multi-Agent Command Center

OpenAI launched Codex in April 2025 as a cloud-based software engineering agent capable of working on many tasks in parallel . By April 2026, it had surpassed 4 million weekly active users . The platform has since expanded into a family of models and interfaces that span an app, CLI, IDE extensions, and cloud, each reinforcing the others .

Key milestones in Codex's evolution:

• GPT-5.2-Codex (December 2025) — optimized for agentic coding and defensive cybersecurity, with improved long-horizon task handling and stronger Windows environment support .
• GPT-5.3-Codex (February 2026) — 25% faster, combining frontier coding performance with broader reasoning and professional knowledge, positioned as a step toward a full "work on a computer" agent .
• Codex desktop app (February/March 2026) — a "command center" for managing multiple parallel agents on macOS and Windows .
• Codex Security (March 2026) — an application-security agent designed to identify and fix software vulnerabilities .
• GPT-5.5 in Codex (May 2026) — the newest frontier model for complex coding, computer use, knowledge work, and research workflows .

Codex has become production-ready infrastructure that developers describe as fundamentally changing how they build software .

Anthropic: Claude Code — From Research Preview to $2.5 Billion Product

Anthropic's Claude Code, launched at Code with Claude 2025, has achieved the most dramatic commercial trajectory in the AI coding market. It crossed $500 million in annualized run-rate revenue within months of its May 2025 launch, reached $1 billion by late 2025, and exceeded $2.5 billion by February 2026 — a velocity that outpaced even ChatGPT's early adoption . Anthropic's overall revenue grew from roughly $9 billion at the end of 2025 to more than $30 billion by spring 2026, largely driven by Claude Code .

By the numbers, Claude Code's reach is remarkable :

• Accounts for 4% of all GitHub public commits, with a projection to reach 20% or more of daily commits by the end of 2026.
• 80% or more of Anthropic's own engineers use it daily.
• Achieved a 46% "most loved" rating among developers in the Pragmatic Engineer survey (15,000 respondents), well ahead of Cursor (19%) and GitHub Copilot (9%).
• At Code with Claude 2026 (May 6), Anthropic announced Managed Agents and deeper integrations with GitHub, Vercel, Datadog, and Bun .

On May 28, 2026, Anthropic released Claude Opus 4.8 — a hybrid reasoning model with a 1-million-token context window that pushes the frontier for long-running agentic tasks. Anthropic reports it is roughly four times less likely than Opus 4.7 to let flaws in its own code pass unremarked .

How the Engineer's Role Is Changing

The shift from writing code to supervising agents is not a distant prediction — it is the current operating model at the world's largest software organizations. Google's engineers, Pichai explained, are increasingly acting as reviewers and orchestrators rather than line-by-line coders, using AI agents that plan, write, test, and execute complex tasks .

Anthropic's 2026 Agentic Coding Trends Report described the transformation clearly: in 2025, coding agents moved from experimental tools to production systems that ship real features to real customers. AI now handles entire implementation workflows — writing tests, debugging failures, generating documentation, and navigating increasingly complex codebases. The report predicts that single agents will soon become coordinated teams of agents, with tasks that previously took hours or days completed with minimal human intervention .

This redefinition of the engineering role shows up across the major platforms:

• OpenAI's Codex now supports parallel agent execution, letting developers delegate long-running work to multiple agents simultaneously and review their output in a unified interface .
• Anthropic introduced auto mode for Claude Code on March 24, 2026, allowing the agent to autonomously decide which actions are safe to take, with built-in AI safeguards reviewing each action before execution .
• Microsoft's MAI-Code-1-Flash was trained inside GitHub Copilot's production harness — a design choice that means the model learned real agentic coding workflows, not just isolated code generation .

Productivity gains are dramatic. Claude Code demonstrated the ability to architect a complex distributed system in one hour — work that the report claimed had previously taken a Google project a full year . Microsoft claims MAI-Code-1-Flash uses up to 60% fewer tokens on complex tasks compared to comparable models .

The Labor Market: Bifurcation, Not Extinction

In the user's question, specific labor-market numbers were referenced — a 30% increase in U.S. engineering job postings, and a nearly 20% decline in employment for developers aged 22–25. Those precise figures could not be independently verified in the sourced materials. The available evidence, however, paints a consistent picture of a profession undergoing bifurcation rather than collapse.

Rising Demand at the Top

Anthropic's trends report states that companies are hiring more engineers, not fewer, as AI enables faster shipping and higher-value work . The demand is shifting toward senior engineers who can architect systems, review AI-generated output, and make high-level design decisions. At Google, the internal targets for agent-assisted code changes, combined with the company's statement that engineering headcount continues to grow, suggest that AI is being used to amplify output rather than to replace engineers outright .

The Junior Developer Experience Gap

The most urgent concern in the sourced materials is what happens to early-career engineers. Junior developers have traditionally built skills through routine coding tasks — fixing bugs, writing tests, implementing straightforward features. Those exact tasks are the ones now most efficiently absorbed by AI agents. Multiple sources describe this as an "experience gap" problem: if AI handles entry-level coding work, how do new engineers learn to become seniors ?

No source in the provided materials provides a verified solution to this challenge. The implication is that the profession will need new training pipelines, mentorship structures, and a redefined career ladder — but those changes are still being worked out.

What Comes Next

The trajectory is unambiguous. Google went from 25% AI-generated code to 75% in eighteen months. Claude Code went from zero to $2.5 billion in annualized revenue in less than a year. OpenAI's Codex grew from a single-agent CLI to a multi-agent platform spanning desktop, cloud, and IDE in roughly the same period .

The unanswered questions are not about whether AI coding agents will continue to improve — they will — but about how engineering organizations, educational institutions, and individual developers will adapt to a profession where the act of writing code is increasingly handled by machines, and the human role is to steer, review, and decide what to build.