AutoDev: Automated AI-driven development by Microsoft

Original Article ↗ Hacker News Discussion ↗

Scope of AutoDev and Similar Tools

  • Seen as an AI “assistant” that automates parts of existing workflows (tests, refactors, simple bugfixes), not full autonomous engineering yet.
  • Compared to Devin: one is framed more as a “builder/assistant,” the other as a more autonomous “architect,” though both are early and demo-like.
  • Some note similar open-source tools already exist (e.g., IDE-integrated agents, custom DSLs), leading to confusion and calls for clearer naming.

Evolving Role of Software Engineers

  • Many predict a shift from “writing code” to:
    • System and architecture design
    • Requirements gathering and domain modeling
    • Integration, verification, and exploratory testing
    • Project/program management and stakeholder communication
  • Entry- and mid-level “CRUD/plumbing” work is seen as most at risk; niche expertise, legacy systems, and deep debugging may remain safer.

Productivity, Benchmarks, and Limits

  • Heavy users report maybe 2–3x personal productivity at best; far from “100x engineers.”
  • Benchmarks like HumanEval are criticized as leetcode-style, narrow, and possibly contaminated by training data. Real-world relevance is questioned.
  • Some think agents plus LLMs will eventually handle steps 2 and 4–8 of the dev lifecycle (design, coding, testing, troubleshooting), others see major gaps in autonomy and robustness.

Economic and Social Impact

  • Strong concern that companies will use AI to reduce headcount rather than hours, shrinking middle-class dev jobs.
  • Historical analogies: portrait painters vs photography, textile workers vs mechanization, spreadsheets vs accountants. Some see eventual new roles; others fear a harsher transition without safety nets.
  • Debate over whether this is “creative destruction” or a path to mass redundancy, including other white‑collar roles.

Process, Testing, and Requirements

  • Many expect AI-driven dev to center on precise specification: tests, property-based specs, or specialized requirement languages, not plain English.
  • Some see this as offloading “fun coding” to AI and leaving humans with the hard, ambiguous, and sometimes “shitty” requirements work.

Broader AI/AGI Debate and Skepticism

  • Split between those who see LLMs as overhyped “stochastic parrots” and those who view them as early but fast-improving general tools.
  • Concerns about ossification: models may reinforce current dominant tools (e.g., pandas, table-based layouts) and slow deeper innovation.
  • Liability and alignment questions remain unresolved: who is responsible when AI-generated code fails is seen as legally and practically unclear.