<aside> 📌

Summary

• The bottleneck in the AI era has shifted from writing code to verifying whether the code is right.
• Investments divide into What (the team's judgement criteria) and How (workarounds that compensate for the limits of today's tools). The What endures. The How fades as models and tools improve.
• Treat today's compensating investments as disposable. Don't chase perfection.
• Investing in the mechanics of inspection is more than ten times as effective as polishing prompt wording (Stanford HAI).
• The guiding principle: "Don't delegate thinking, delegate toil." </aside>

The Question

<aside> 💭

Coding agents have come on in leaps and bounds. Teams and individuals alike have sharpened their skills and prompts, built out their harnesses, and automated a great deal of their work. Meanwhile, the underlying models, agents and tools are advancing at a furious pace.

Are we pouring resources into problems that future models will simply solve for us?

This document sets out the criteria for what to invest in and what to say no to. It draws on a cross-sectional reading of more than forty primary sources from the industry.

</aside>

Terminology

The main concepts used throughout this piece.

• Click to expand

Premise: The Bottleneck Has Moved from Generation to Verification

The cost of writing code has dropped dramatically. But that isn't the solution to the problem — it is the relocation of the bottleneck.

"Code generation was never the bottleneck of software engineering. The real bottleneck is verification." — InfoWorld

The data bears this out.

• CodeRabbit 2026 analysis: merged PRs are up 98% and PR sizes are up 154%. Yet 61% of developers say "AI output looks right but can't be trusted".
• QA Trends (Tricentis 2026): at least 60% of AI-generated code contains issues requiring intervention.
• MSR 2026 (analysis of 806 repositories): after AI adoption, code complexity is up 41%, static-analysis warnings up 30%, and change failure rate up 30%.

Implication: AI-driven productivity, unaccompanied by investment in verification infrastructure, becomes mass-production of technical debt.

The Criteria: Investments That Endure, Investments That Fade