Verification as the AI coding bottleneck
Simon Willison: AI coding shifts the bottleneck from writing to verification Microsoft DeveloperTL;DW
- Genetic engineering—using AI agents to write and test code—only became reliably productive in November 2025 with Claude Opus 4.5 and GPT-5.1, after both labs spent all of 2025 doing RL on coding tasks.
- Strong DM's software factory used scenario-based testing (agents running edge cases), digital twin clones of external APIs (Slack, Okta, Jira) to remove rate limits, and autonomous verification—no human code review required.
- Active refactoring during review is anti-pattern for humans but works great for agents: nitpick variable names, remove duplication, request explanations, improve comments without risk of offending or delaying feedback.
- Prototype early and often using agents; save prototypes as input files for later work. This shifts developer role from typing code to architecture and design, especially for API surface validation.
- Continuous deployment to preview environments for every PR reduces code review risk to near-zero; use agents to build GitHub Actions workflows with 'ask clarifying questions first' to avoid hallucinations.
- Reduce blast radius for mistakes: content security policy headers, sandboxed iframes, and WebAssembly (WASI on servers) let you run untrusted generated code safely without risk of file access or network exfiltration.
- Zero tolerance for flaky tests; agents excel at reproducing race conditions and environmental bugs (like Python 3.4 on Linux in CI) that would take humans hours to debug manually.
- Documentation trust is fragile: use agents to keep docs in sync with code changes via `git diff` prompts, but strip all promotional language, rationales, and opinions that make docs feel artificial and misleading.
- Agentic documentation works because agents produce uncreative, accurate, short prose—exactly what code docs need. Keep a rule: no opinions, no rationales, or readers will stop trusting your docs forever.
- Conceptual integrity matters in 2026 as much as 1978: generating features via prompt alone creates shapeless products you can't explain or maintain. Plan features in markdown, maintain consistent mental models of your system.
TL;DW
- Genetic engineering—using AI agents to write and test code—only became reliably productive in November 2025 with Claude Opus 4.5 and GPT-5.1, after both labs spent all of 2025 doing RL on coding tasks.
- Strong DM's software factory used scenario-based testing (agents running edge cases), digital twin clones of external APIs (Slack, Okta, Jira) to remove rate limits, and autonomous verification—no human code review required.
- Active refactoring during review is anti-pattern for humans but works great for agents: nitpick variable names, remove duplication, request explanations, improve comments without risk of offending or delaying feedback.
- Prototype early and often using agents; save prototypes as input files for later work. This shifts developer role from typing code to architecture and design, especially for API surface validation.
- Continuous deployment to preview environments for every PR reduces code review risk to near-zero; use agents to build GitHub Actions workflows with 'ask clarifying questions first' to avoid hallucinations.
- Reduce blast radius for mistakes: content security policy headers, sandboxed iframes, and WebAssembly (WASI on servers) let you run untrusted generated code safely without risk of file access or network exfiltration.
- Zero tolerance for flaky tests; agents excel at reproducing race conditions and environmental bugs (like Python 3.4 on Linux in CI) that would take humans hours to debug manually.
- Documentation trust is fragile: use agents to keep docs in sync with code changes via `git diff` prompts, but strip all promotional language, rationales, and opinions that make docs feel artificial and misleading.
- Agentic documentation works because agents produce uncreative, accurate, short prose—exactly what code docs need. Keep a rule: no opinions, no rationales, or readers will stop trusting your docs forever.
- Conceptual integrity matters in 2026 as much as 1978: generating features via prompt alone creates shapeless products you can't explain or maintain. Plan features in markdown, maintain consistent mental models of your system.
Willison argues that as agents generate hundreds of lines daily, quality depends on comprehensive test suites, sandboxing, and reducing blast radius—not just reviewing diffs. Covers scenario-based testing, prototyping against new APIs, flaky-test elimination, and why accumulated prompts without conceptual oversight produce architecturally incoherent systems.