Claude Code is suddenly everywhere inside Microsoft

Original Article ↗ Hacker News Discussion ↗

Claude Code vs Copilot and other agents

  • Many developers report Claude Code as “just better” than GitHub Copilot, especially for larger refactors, multi-file changes, and long-running tasks.
  • Several use Copilot only as a gateway to Anthropic models (Sonnet/Opus) via Copilot CLI or OpenCode, bypassing Microsoft’s own agent UX.
  • The Claude Code CLI-first, repo-aware workflow is widely praised; Copilot’s VS Code and IntelliJ integrations are often called sluggish, brittle, or unintuitive.
  • Some say Copilot CLI is now “good enough” and close to Claude Code when configured well, but others still find it noticeably weaker.

Microsoft’s AI strategy and the “1 engineer, 1 month, 1M LOC” flap

  • A LinkedIn post by a senior Microsoft engineer about “1 engineer, 1 month, 1 million lines of code” and rewriting C/C++ to Rust via AI triggered strong backlash.
  • Debate over whether this is a personal research “North Star” or indicative of broader corporate goals; some see the later “clarification” as damage control.
  • Almost everyone agrees LOC as a productivity metric is absurd and dangerous, especially when supercharged by LLMs.

Perceived product decline and AI “slop”

  • Many tie worsening Windows reliability, broken sleep/standby, and erratic updates to AI-driven development and misaligned incentives.
  • Some argue Microsoft prioritizes shipping features and AI branding over quality; engineers confirm internal incentives reward shipping, compliance, and AI, not bugfixing.
  • Multiple commenters report abandoning Windows for Linux/macOS because of enshitification and Copilot/Recall-style features.

Naming, branding, and confusion around Copilot

  • Widespread frustration that “Copilot” now labels many unrelated products: Windows chat, GitHub tools, Office/M365 features, Azure, Xbox, etc.
  • This causes constant miscommunication: criticism of one Copilot variant is often answered with praise for a different one.
  • Microsoft’s long history of chaotic naming (.NET, Live, One, 365, Xbox variants) is heavily mocked.

Models vs harnesses: Opus, Codex, Gemini

  • Some say Anthropic’s Opus 4.5 is currently the best for agentic coding; others claim GPT‑5.2 Codex produces better raw code but is hampered by weaker harnesses (e.g., Codex CLI, OpenCode).
  • Gemini gets mixed reviews: some find Gemini 3 Flash/Pro extremely cost-effective and competitive, others call it hallucination-prone or “lazy” as an agent.
  • A recurring theme: the harness/agent UX (Claude Code, Copilot CLI, Antigravity, Codex CLI) matters as much as the underlying model.

Internal culture and dogfooding

  • Multiple anecdotes say Microsoft and Apple engineers heavily use Claude Code internally, often on macOS, rather than Microsoft’s own AI tools.
  • Commenters see this as evidence both of Claude’s quality and of Microsoft’s failure to dogfood and harden Copilot-based workflows.

Security, privacy, and AI-generated future

  • Concerns raised about sensitive code and credentials flowing through LLMs; some suggest architectures where secrets never enter the model context.
  • Several predict most software will eventually be majority AI-generated, raising questions about bloat, maintainability, and how to measure “better code” once “more code” is trivial.