Codex for almost everything

Original Article ↗ Hacker News Discussion ↗

Scope of the Codex Update

  • Many find the announcement vague: unclear whether “major update” means a new underlying model, new desktop features, or mainly UI and tool changes.
  • Several note that most capabilities (coding agents, computer use, plugins) already exist in competing tools, particularly Claude Cowork/Code; Codex is seen as catching up and iterating on UX rather than pioneering.

Computer Control & Agents

  • Strong split:
    • Enthusiasts want full “Star Trek”‑style agents that plan trips, manage files, fix servers, test apps, manipulate browsers, etc. Some share striking success stories (e.g., fixing complex Linux issues, planning vacations, automating social media campaigns, scheduling, sysadmin work, even taxes).
    • Skeptics see giving an LLM broad OS control as “a nightmare waiting to happen,” insisting on sandboxes, separate machines, or minimal permissions.
  • Background GUI use (hidden cursor operating apps in parallel) is viewed as a powerful but risky new surface.

Security, Privacy, and Sandboxing

  • Serious concern that Codex and similar tools read sensitive files (e.g., API keys, personal data) without sufficiently clear prompts or consent.
  • Linked examples of agents exfiltrating secrets or modifying tests instead of code reinforce fears.
  • Agreement that non‑technical users tend to blindly approve permissions, increasing risk.
  • Some predict this will drive more restrictive OS designs, mandatory cloud storage, and more surveillance.

Target Users, UX, and “Coding Without Code”

  • Debate over who really wants this:
    • Many argue non‑technical workers want a simple “one button” interface and will accept agents if they reliably get work done.
    • Others claim “ordinary people” distrust or dislike AI and won’t tolerate opaque, constantly changing AI‑generated UIs.
  • Ongoing argument about “hiding the code”:
    • Some welcome a world where non‑coders “vibe code” via agents and never see source, enabling lots of DIY, non‑production tools.
    • Others insist code remains a critical artifact for correctness, maintainability, and security, and that AI‑generated slop will create massive technical and societal debt.

Impact on Software Engineering

  • Many programmers use Codex/Claude as pair‑programmers: they design structure, tests, and architecture, and let the model fill in boilerplate.
  • View that LLMs are great at CRUD, glue code, config, and log analysis, but still require expert supervision for complex design and verification.
  • Some predict traditional IDEs and “code quality” norms will erode as agents become the primary interface; others argue complexity and performance guarantees will keep disciplined engineering relevant.

Model Quality, Limits, and Business Dynamics

  • Mixed experiences comparing Codex (GPT‑5.4) vs Claude Opus/Sonnet: some find Codex clearly better, others say Claude is “night and day” superior for data and agentic tasks.
  • Heavy frustration with rapidly changing rate limits and pricing on both sides; perception that generous limits are temporary promos.
  • Discussion that both labs are subsidizing usage to gain market share; concern this is anti‑competitive and sets up future “rug pulls.”

Platform & Open Source Concerns

  • Annoyance that full computer‑use is Mac‑only; Windows support is missing and Linux is largely sidelined to CLI tools, despite some third‑party efforts.
  • Some hope open‑weight models and local agents will eventually catch up, but many report current open models still lag behind frontier models for complex tasks.

Broader Social & Ethical Concerns

  • Worries about enabling scammers, mass low‑quality “slop” software, and large‑scale surveillance or state use of AI.
  • Fear of workers training their own replacements, though others argue verification, liability, and office politics will delay or limit full automation.