Source note
AI demands more engineering discipline. Not less
Summary
The piece argues that AI coding increases the need for engineering discipline because code is now cheap to generate and costly to validate. It points teams toward specifications, observability, production feedback, and repeatable validation as the durable assets.
Problem
- AI coding tools can generate common code faster and cheaper than before, so human review of lines of code alone becomes a weak control.
- Teams often store product intent, edge cases, and operational knowledge inside old code and people’s heads; that makes rewrites and migrations risky.
- Nondeterministic AI systems need tighter feedback loops because users still need stable behavior, durable data, and reliable transactions.
Approach
- Treat source code as a disposable output of shared understanding, similar to immutable infrastructure replacing mutable servers.
- Move rigor into artifacts that can be checked and replayed: specs, invariants, behavioral tests, characterization tests, capture/replay setups, traffic splitting, traces, and production evals.
- Use agentic LLM harnesses with tools, function calling, and MCP-style integration to generate or regenerate code, then validate it against encoded behavior.
- Review higher-level artifacts such as architecture diagrams, requirements, and validation signals instead of relying only on line-by-line code review.
Results
- The piece claims that by November 2025, Anthropic Opus 4.5 made AI-generated code roughly as good as a median software engineer for common patterns, with lower time and cost; it gives no benchmark dataset or score.
- It dates the rise of practical agentic coding harnesses to mid-2025, with precursors in late 2024 and broad usability by the end of 2025.
- It estimates that only about 5% of software teams, and less than 10%, work in short feedback loops today.
- Honeycomb issued an internal AI mandate in August 2025, used as a concrete adoption marker rather than a measured experiment.
- The essay gives no controlled quantitative results, datasets, or baseline comparisons; its strongest concrete claim is that code production economics changed faster than validation practices.