The software development industry is grappling with an unexpected reality: artificial intelligence coding tools, while dramatically improving individual developer productivity, have failed to deliver proportional gains in overall project delivery speed. This finding, detailed in a recent analysis by Agoda, challenges conventional assumptions about where bottlenecks exist in the software development lifecycle.

Leonardo Stern, a software engineer at Agoda, frames this phenomenon as a rediscovery of principles outlined decades ago by Fred Brooks in his seminal work 'No Silver Bullet.' Brooks argued that improvements targeting only one aspect of the development process produce diminishing returns for overall delivery performance. The current AI coding tool landscape appears to validate this thesis in real-world conditions.

Empirical evidence supports this counterintuitive observation. Research conducted by Faros AI, analyzing telemetry data from over 10,000 developers across 1,255 teams, revealed that organizations with high AI adoption rates completed 21% more individual tasks and merged 98% more pull requests. However, these same teams experienced a 91% increase in pull request review time, indicating that acceleration in one area created new pressure points elsewhere in the development pipeline.

According to Stern's analysis, the bottleneck has shifted upstream to specification and verification phases, areas that fundamentally require human judgment and collaborative decision-making. This shift carries significant implications for engineering team structure and workflow design. Traditional rationale for small, focused engineering teams was partially built on the assumption that coding represented the most significant value-creating activity, with communication viewed as overhead that impeded productivity.

If the highest-value work transitions to collaborative specification and architectural alignment, this logic inverts completely. Communication transforms from a cost to minimize into the primary work itself. Smaller teams maintain advantages not because they reduce coordination overhead, but because they achieve shared understanding more efficiently. Five engineers can genuinely align around intent and edge cases in ways that fifteen typically cannot.

To navigate this new landscape, Stern introduces a three-stance taxonomy for engineers working with AI-generated code. The 'white box' model requires humans to read and review every line of generated code, an approach that fails to scale when AI agents can produce thousands of lines per hour. The 'black box' or 'vibe coding' approach involves shipping AI-generated output with minimal verification, which proves too risky for production systems serving large user bases.

Stern advocates for a middle path he terms the 'grey box' approach, which maintains human accountability at two critical junctures: writing specifications precise enough for AI agents to execute correctly, and verifying results against evidence rather than inspecting implementation line by line. This methodology keeps engineers fully responsible for what ships to production while leveraging AI capabilities effectively.

This reframing of the review process from code inspection to evidence evaluation reflects broader industry trends toward spec-driven development. In this paradigm, specifications become the executable source of truth for systems, with generated code treated as a downstream, regenerable artifact. Both perspectives converge on the observation that human authority is migrating upward in the abstraction stack, from writing code to defining and governing intent.

The implications extend beyond individual workflow optimization to fundamental questions about software engineering's future. Rather than replacing human developers, AI tools appear to be elevating their role to focus on higher-level design, specification, and verification tasks that require human judgment and collaborative problem-solving. This evolution suggests that the most successful engineering organizations will be those that restructure their processes to optimize for specification quality and verification effectiveness, rather than raw coding output.

As the industry continues to integrate AI coding assistants, understanding these dynamics becomes crucial for engineering leaders seeking to realize genuine productivity gains rather than simply shifting bottlenecks to different parts of their development pipeline.