What it feels like to work with Mythos

Original Article ↗ Hacker News Discussion ↗

A glowing review of Anthropic’s new Mythos/Fable AI coding agent prompts sharp debate over how transformative these large language models really are. Commenters acknowledge impressive feats like multi‑hour autonomous coding sprints and complex data visualizations, but point to buggy demos, unmaintainable code, high token costs and heavy guardrails as evidence that human oversight and careful architecture remain essential. The thread also surfaces broader concerns about hype, job displacement, shifting notions of “code quality,” and whether current gains are incremental or a genuine step change in software development.

Overall sentiment

  • Thread is sharply split between enthusiasm and skepticism.
  • Supporters see Mythos/Fable as a genuine capability jump for complex coding and analysis.
  • Critics see overhyped marketing, “vibes-based” evaluation, and thin technical evidence from non-engineers.

Capabilities & use cases

  • Some users report clear wins vs earlier models (Opus 4.8, GPT‑5.5, Qwen, DeepSeek) on:
    • Deep code review and refactors in large projects.
    • Complex performance work (e.g., a Rust Lua interpreter).
    • Building substantial web apps and tools from specs.
    • Systematizing prompt guidelines and “agent” configurations.
  • Others find it only incrementally better, with familiar issues: hallucinations, overtalking, ignoring constraints, getting stuck in loops.

Long-running agents & harness

  • The 9.5‑hour “Concord” build provokes debate:
    • Pro: no human dev could deliver that much from a 19‑page spec in a day.
    • Con: industry wants latency in seconds; long agent runs often drift and need rollback.
  • Several argue most “magic” comes from the harness: teams of sub‑agents, tooling, and good project structure, not just the base model.

Code quality, correctness, and maintainability

  • Many engineers focus on missing details: tests, security, architecture, extensibility, and cost of future changes.
  • Reported issues:
    • Isochrone map has serious factual and UI errors.
    • Games and demos are buggy or break after a few steps.
    • Example repo code called “slop” / “unmaintainable.”
  • Strong concern about the article’s hand‑wave that “a software engineer will iron out the remaining bugs.”
  • Ongoing debate:
    • One side: if the behavior is good and models can continually refactor, internal code quality matters less.
    • Other side: complexity, silent corruption, and compounding “oopsies” still make this unsustainable without strong human design and verification.

Safety, guardrails, and censorship

  • Fable’s aggressive cybersecurity/bio guardrails frequently block exactly the code‑review work people want, forcing a fallback to weaker models.
  • Some report “gaslighting” and silent self‑corruption when the model decides a task is unsafe.

Economics, ROI, and access

  • Users note high token burn: single sessions consuming large chunks of weekly quotas; fears of being “priced out” after promo periods.
  • Debate over whether automation is really cheaper than humans at current prices and quality; calls for concrete cost‑per‑deliverable numbers, which the article omits.

Impact on developers & work

  • Some devs feel 2–3× more productive and see strong ROI; others have already reduced LLM usage due to quality and outage risks.
  • Broad agreement that:
    • Models are powerful for low‑stakes, short‑lived or side projects.
    • High‑stakes, long‑lived systems still need significant human architecture, domain understanding, and review.