Manoj Agarwal, a veteran developer who has built systems at Microsoft, Twitter, and Stripe, says the hardest part of his job now isn't writing code—it's reviewing the flood of AI-generated code that his tools produce. And he's not alone: a 2025 Stack Overflow survey of over 65,000 developers found that while 84% are using or plan to use AI coding assistants, nearly half (46%) actively distrust the accuracy of these tools, and 66% find that AI output that is "almost right" is their biggest frustration.

AI Coding Tools Have Become Mainstream, and They're Getting More Powerful

In a world where GitHub Copilot can suggest entire functions and Claude Code can autonomously edit files, run tests, and open pull requests, software development is accelerating at breakneck speed. According to Agarwal's account, originally reported by Jawlah, the once-clear line between human-written and AI-generated code is blurring inside the tools that Windows developers rely on every day—Visual Studio and Visual Studio Code.

The numbers are staggering. Stack Overflow's 2025 developer survey reveals that 51% of professional developers now use AI tools daily. Agarwal himself spends roughly $100 a month on a suite of products including Claude Code, Microsoft Copilot, and Lovable. "A project that needed ten engineers two years ago may only need half that number today," he told Jawlah, though he frames that estimate as personal observation rather than an industry benchmark.

But the acceleration has a shadow. The same survey found that 45% of developers say debugging AI-generated code is more time-consuming than writing it from scratch, and 46% actively distrust the output they receive. The paradox is sharp: AI can produce more code in an hour than a human could in a day, but that code often contains subtle flaws that take even longer to unearth.

For Developers, It's a Review Deluge. For Users, It's a Mixed Bag.

If you're a Windows developer, the impact is immediate. Your daily workflow now likely includes reviewing not only your teammates' work but also large chunks of AI-authored suggestions that look polished but may harbor logic errors, security gaps, or licensing violations. Agarwal describes the "accumulation of responsibilities for reviewing AI-generated outputs" as a direct contributor to burnout.

For everyday Windows users and IT administrators, the software you depend on—from line-of-business apps to system utilities—is increasingly shaped by these tools. Faster feature updates may come with more frequent patch releases. A flaw in an AI-generated PowerShell script or a misconfigured infrastructure-as-code template can cascade into production outages. As GitHub's own documentation notes, AI reviews should never be treated as comprehensive validations; human oversight remains critical.

The burden is especially heavy on junior engineers. Agarwal warns that entry-level positions, which traditionally serve as both a training ground and a source of routine code contributions, are shrinking. "Most available jobs currently go to those with enough experience to effectively use AI tools," he told Jawlah. Organizations risk losing the pipeline that produces the next generation of senior architects.

From Autocomplete to Autonomous Agents: A Timeline of AI Coding

AI assistance in coding didn't materialize overnight. GitHub Copilot launched in 2021 as a smarter autocomplete, offering line and function completions powered by OpenAI's Codex model. By 2024, tools like Claude Code had evolved into "agentic" systems capable of reading entire codebases, modifying multiple files, running terminal commands, and integrating with version control. The jump was profound: developers went from accepting a suggestion to delegating multi-file refactors.

Windows developers were early beneficiaries. The deep integration of Copilot into Visual Studio and VS Code meant that C#, PowerShell, Python, and TypeScript programmers could summon AI assistance without switching contexts. GitHub's documentation highlights that Copilot can suggest everything from a single line to entire boilerplate classes, across a broad range of languages commonly used in the Windows ecosystem.

But the foundational technology hasn't solved its core problem: large language models generate plausible text, not necessarily correct software. A missing null check, an outdated API call, or an overly permissive access control rule can lurk inside syntactically perfect code. Experienced engineers who have wrestled with production outages know that elegant-looking mistakes are far more dangerous than obvious failures.

How to Keep Up Without Burning Out: Practical Steps for Developers and Teams

The solution isn't to abandon AI—it's to build workflows that scale verification alongside generation. Based on current best practices from GitHub's engineering teams and expert consensus, here are concrete steps:

  • Write acceptance criteria first. Before prompting an AI, define the intended behavior, edge cases, security constraints, and compatibility requirements. Vague prompts produce vague code.
  • Review the actual diff, not just the summary. AI summaries are useful but can gloss over critical details. Inspect authorization checks, error handling, and dependency changes line by line.
  • Keep changes small and scoped. A single pull request that touches dozens of files is a review nightmare. Break AI-generated work into narrow, well-defined pieces.
  • Assign human ownership. The person merging the code must be able to explain its design and accept responsibility for its behavior in production.
  • Treat prompts as engineering artifacts. For high-stakes work, document the exact instructions given to the AI so the rationale can be reproduced and audited.
  • Design apprenticeship programs for the AI era. Instead of eliminating junior roles, teach new developers to use AI critically: read code before generating it, debug with explanations, and own small features from inception to post-release support.

For Windows IT administrators managing development teams, this means investing in static analysis tools, automated security scans, and review environments that go beyond quick browser-based inspections. Tools like Microsoft Defender for DevOps and built-in secret scanning in GitHub can catch many problems, but no automated check replaces a human who understands the organization's authentication model and data boundaries.

The Future: Senior Engineers Become Even More Critical

Agarwal is firm that "AI still needs the seasoned engineer," emphasizing that these models lack the holistic vision required to design a system. His advice to peers: don't fear AI, but dive in and learn how to use it efficiently—while always remembering that the efficiency gain comes with an obligation to verify.

The upshot for the Windows software world is twofold. First, expect the velocity of updates and new features to increase, but also expect that quality assurance will become more contested. Second, the market value of experienced developers who can navigate both the technical and business sides of a system will rise sharply. The ideal engineer of the near future won't be the fastest typist but the person who can translate messy business needs into testable requirements, evaluate AI output against architectural constraints, and ensure that code—regardless of its origin—earns the right to run.