Is it all just vapourware?

Original Article ↗ Hacker News Discussion ↗

Claims that “agentic” AI coding tools are mostly vapourware collide with reports from developers who say state-of-the-art models meaningfully accelerate boilerplate work, debugging, and performance tuning. Many agree the underlying tech is real but argue valuations, hype, and fully autonomous “software factories” are overblown, with LLMs often generating fragile, hard-to-maintain code and creating extra review and integration overhead. The exchange highlights a widening gap between lab demos and long-term, production-quality software, and raises questions about where genuine productivity gains are appearing versus where AI is just amplifying noise and technical debt.

Reality of the Tech vs. Valuation Hype

  • Broad agreement that LLMs are real and often impressive, especially for coding, research help, and some military/industrial uses.
  • Many see current valuations and grand “agents will do everything” narratives as bubble-like or pyramid-scheme-adjacent (with GPU vendors and labs at the top).
  • Critique: products are being launched “before the value is there,” and much of the market is driven by hype, jargon, and investor pressure.

Agentic Coding in Practice

  • Some report fully “agentic” loops that can pick tasks, write code, run tests, and ship features, especially in small shops and greenfield projects.
  • Others find agents thrash, waste credits, generate huge low-quality PRs, and require so much supervision that they slow things down.
  • There’s concern that agents encourage endless refactoring and side projects rather than finishing real, scoped work.

Productivity Claims and Evidence Gap

  • A camp claims multiple “jaw-dropping” moments and 5–10x productivity on real products, especially for boilerplate, performance tuning, refactors, and cross-cutting changes.
  • Another camp says nothing actually got faster once review, debugging, and integration are counted; downstream bugs and noise increase.
  • Several note that organizational bottlenecks (meetings, misaligned priorities) can easily absorb any raw speed gains.

Open Source Experience

  • Argument that if 10x gains were real, major OSS projects would show a dramatic uptick in features and velocity; many observers don’t see this.
  • Maintainers report high volumes of low-quality LLM PRs and are increasingly banning or discouraging AI-generated contributions.
  • Counterpoints:
    • Good AI-assisted code may be indistinguishable from human code.
    • Maintainer time, not code generation, is the bottleneck; bans are often about review load, not pure quality.
    • Some concrete high-impact AI-assisted optimizations in major projects are cited.

Best-Use Patterns vs. Failure Modes

  • Stronger consensus on “sweet spots”:
    • Boilerplate and rote code.
    • Debugging and tracing issues.
    • Performance profiling and micro-optimizations.
    • Writing tests, fuzzers, and instrumentation.
    • Prototyping and “disposable” tools or internal apps.
  • Known failure modes:
    • Complex, long-lived architectures and safety-critical systems.
    • Agents trying to plan and implement whole features autonomously.
    • Deep domain research (e.g., advanced physics) where answers are confidently wrong.

Role of Expertise and Workflow Design

  • Many say LLMs strongly amplify existing expertise: great in the hands of strong engineers, dangerous in the hands of novices.
  • Comparisons are made to power tools: high leverage, no safety guards.
  • Effective use often means tight human-in-the-loop cycles, strong testing (including fuzzing and visual regression), clear specs, and constrained tasks.

Economics, Organizations, and Incentives

  • Some argue software jobs are already performative in many companies; AI excels at “the illusion of progress,” which may be all that’s required for stock and career incentives.
  • Others emphasize that real, non-VC-funded businesses are already getting tangible value from agentic workflows and monitoring infrastructure.
  • Concern that churn, abandonware, and “MVP culture” are degrading user experience as teams ship AI-powered products without enough testing or polish.

Long-Term Concerns and Hype Cycle

  • Noted “Cambrian explosion” of tools, many quickly abandoned; heavy churn makes learning stable workflows hard.
  • Some see growing backlash and dissatisfaction as users confront limitations and integration pain.
  • Several expect a shakeout: one-shot “magic agent” products will fade, but more grounded engineering-focused uses of LLMs will remain and mature.