AI Coding Assistants Haven't Accelerated Software Delivery Because Coding Was Never the Real Bottleneck
Key Takeaways
- ▸AI coding assistants have increased individual developer output but not proportionally accelerated overall project delivery, indicating coding was never the primary bottleneck
- ▸The real bottleneck has shifted to specification and verification work, which require human judgment and cannot be easily automated
- ▸Teams should adopt a 'grey box' verification model: high-fidelity specifications written by humans with AI handling implementation, then evidence-based review rather than line-by-line code inspection
Summary
Agoda engineer Leonardo Stern has published an analysis arguing that while AI coding assistants have measurably increased individual developer productivity, project-level delivery velocity improvements remain surprisingly modest. The core insight is that coding was never the actual bottleneck in software development—specification and verification, which require human judgment, are. This observation aligns with industry research from Faros AI showing that teams with high AI adoption completed 21% more tasks and merged 98% more pull requests, yet PR review time increased by 91%, suggesting productivity gains at the coding stage have shifted bottlenecks upstream.
Stern frames this as a rediscovery of Fred Brooks' classic "No Silver Bullet" argument that improvements in one part of the development lifecycle produce diminishing returns when other areas remain constrained. The implications extend beyond tool adoption to how engineering teams should be structured. If specification and verification are now the highest-value activities, then communication and shared understanding become the work itself rather than overhead to minimize, favoring smaller, more tightly-aligned teams.
Stern proposes a "grey box" approach to AI-generated code where humans maintain accountability at two critical points: writing precise specifications for AI agents and verifying results against evidence rather than inspecting implementation line-by-line. This approach keeps human authority focused on defining intent and governing architecture, with implementation increasingly delegated to AI. The analysis aligns with emerging industry consensus on spec-driven development, where specifications become the executable source of truth and generated code is treated as a regenerable artifact.
- Software engineering's value-creation center has migrated from writing code to defining intent and governing architecture, with implications for team structure and organizational design
- Industry data shows high-AI-adoption teams complete more tasks but experience 91% longer PR review times, confirming productivity gains shifted rather than eliminated bottlenecks
Editorial Opinion
Stern's analysis provides a sobering but important reality check for the AI-in-software-development narrative. While the finding that coding wasn't the real bottleneck may disappoint those expecting dramatic velocity multipliers, it offers more valuable insight: the future of engineering management lies in specification discipline, architectural clarity, and verification rigor—skills that have been chronically underinvested in many organizations. The reframing of code review from inspection to evidence evaluation is particularly pragmatic and points toward how engineering teams can truly leverage AI without creating new vulnerabilities or shifting all risk downstream.
