What's the largest software project AI can complete on its own?

Original Article ↗ Hacker News Discussion ↗

Developers are probing how far current AI models can go in autonomously building software, from shells, compilers and retrieval engines to full apps and even OS-like environments. Many report that while AI can generate large amounts of working code and clone well-specified systems when guided by strong tests and human-designed architecture, it quickly drifts into brittle, duplicated, or incoherent designs if left unsupervised. The emerging consensus is that AI is most effective as a powerful junior engineer within human-led workflows, and that the harder open problem is long-term maintenance, architectural coherence, and requirements definition rather than raw code generation.

Scope of “AI-complete” Software

  • Several comments argue that “largest project” is the wrong question; the harder, more interesting one is how large a system an AI can maintain over time while preserving coherent architecture, performance, and safety.
  • Others note that reproducing existing software with clear specs/tests is much easier than designing new products with fuzzy requirements.

Real-World AI Project Experiences

  • Reported AI-heavy projects range from:
    • A Bash clone in Rust with thousands of tests and hundreds of issues left for agents to fix.
    • A 180k-LOC investigative retrieval engine with custom WAL, hybrid retrieval, ACLs, and ~3,000 tests, evaluated by government agencies.
    • A full Linux distro packaged as a Mac app built over ~2 weeks.
    • Portfolio management tools, FHIR/SNOMED prototypes, caching proxies, game clones, compiler experiments, and a multi-language codebase with ~2M+ LOC.
  • Typical pattern: AI does most implementation; humans handle architecture, specs, and reviews.

Architecture, Messiness, and Long-Term Health

  • Many report that unsupervised or “vibe-coded” projects quickly devolve into spaghetti:
    • Shortest-path fixes, duct-tape patches, duplication, load-bearing hacks.
    • AI rarely steps back to do big architectural refactors on its own.
  • Some say this resembles junior engineers needing direction; others emphasize humans can autonomously design and maintain complex systems in ways current models cannot.

Testing, Harnesses, and Guardrails

  • Strong consensus that success hinges on:
    • Extensive unit/integration tests, invariants, and end-to-end comparisons with reference implementations.
    • Incremental, small-scope tasks plus continuous human review.
    • Harnesses/loops that allow agents to run many iterations under strict test suites.
  • Disagreement over whether “lots of tests” is sufficient without humans deeply understanding what is being tested.

Capabilities, Limits, and Use Cases

  • Models perform well on:
    • Small to medium components (e.g., ~500-line React components, endpoints).
    • Well-specified ports/clones with rich existing tests.
  • They struggle with:
    • Novel architectures, large unique domains (e.g., CAD kernels), and global consistency.
    • Fully autonomous operation without periodic human course correction.

Benchmark and Data Concerns

  • Some question benchmarks that only cover text-based programs and potentially rely on code seen during training.
  • Several stress that benchmarks with trivial budgets are unrealistic; serious evaluations must allow large search spaces and long runs.