Execution is no longer the dominant constraint

Claim

For most software-driven organizations, the cost of producing artifacts (code, designs, drafts) is falling faster than the cost of deciding what should exist and how it should behave.

Rationale

AI systems dramatically reduce realization latency, enabling rapid generation and iteration. This amplifies the cost of poor framing and weak constraints.

Supporting Signals

• Rapid artifact generation with minimal marginal cost
• Increased verification and correction overhead in mature systems
• Shift in effort toward review, alignment, and validation

Challenges

• Expert slowdowns due to verification tax
• Tool brittleness in complex, stateful systems

Evidence

• The Mythical Man-Month — Foundational software engineering text. Book — requires manual capture of relevant excerpts.
• Code Complete — Comprehensive software construction guide. Book — requires manual capture of relevant excerpts.
• The Impact of AI on Developer Productivity: Evidence from GitHub Copilot — Randomized controlled trial. Captured from Microsoft Research publication page.
• Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity — Rigorous RCT. Web page — requires manual capture (403 on automated fetch).
• Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (arXiv) — Academic preprint with full methodology. PDF — requires manual download.
• AI | 2025 Stack Overflow Developer Survey — Annual survey of 65,000+ developers. Web page — requires manual capture (403 on automated fetch).
• Developer Experience Report 2025 — Industry report on developer productivity. Web page — requires manual capture (403 on automated fetch).
• The State of Developer Experience in 2025 — Full research PDF companion. Gated PDF — requires manual download and capture.
• Is AI Creating a New Code Review Bottleneck for Senior Engineers? — Industry reporting. Web page — requires manual capture (403 on automated fetch).
• Generative AI, Productivity and the Future of Work — Regional Fed analysis. Emphasizes realized benefits depend on redesigning work so human effort moves into higher-order decision-making rather than adding more AI-mediated activity. Supports task reallocation > job loss framing.
• How much does AI impact development speed? An enterprise-based randomized controlled trial — Enterprise-style RCT with 96 full-time Google SEs. Best estimate 21% reduced time-on-task with wide CI. More realistic than toy-task studies. Quality not fully captured in metrics.
• The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers — Three randomized field experiments across real companies. Pooled estimate: +26.08% completed tasks. Bigger gains for less experienced developers. 4,867 developers total. Strongest multi-company field evidence to date.
• Measuring Program Comprehension: A Large-Scale Field Study with Professionals — IEEE TSE. 78 developers across 7 projects, 3,148 hours of telemetry. Found developers spend 58% of time on program comprehension. Pre-AI but directly supports the claim that execution (writing code) was never the dominant time sink.
• Meta-Harness: End-to-End Optimization of Model Harnesses — Stanford/MIT/KRAFTON. Harness optimization outperforms hand-engineering: +7.7pt text classification with 75% fewer context tokens. Proves orchestration (what to store/retrieve/show) dominates raw execution.
• The Impact of Large Language Models on Open-Source Innovation: Evidence from GitHub Copilot — Causal natural experiment (Python supported by Copilot, R not): Copilot availability increases open-source contributions by 28-40%, with incremental contributions rising significantly more than substantive ones. Strongest on-mechanism evidence that execution cost is falling, concentrating value on deciding what to build.
• Debt Behind the AI Boom — Analysis of 304k AI commits: more than 15% of commits from every AI assistant introduce at least one issue, and 24.2% of those survive to HEAD. Mechanism-adjacent support that cheaper execution migrates cost to verification and maintenance.
• Acceleration Whiplash: AI's Effect on Software Delivery — Telemetry across 22k developers: median review time 5X, incidents per PR 3X, +66.2% epics alongside +861% churn. Execution volume rises while verification becomes the binding constraint (31.3% more PRs merging unreviewed), evidencing cost migration to review and authority concentrating where verification is defined.
• Constraint Decay: How Specification Pressure Degrades LLM Coding Performance — Capable configurations lose 30 points on average in assertion pass rates from baseline to fully specified tasks, showing execution-cheap weakens in convention-heavy production. Scope-bound challenge: re-scopable to greenfield vs constrained production, so no demotion.
• 2026 Agentic Coding Trends — Reports developers use AI in roughly 60% of work but fully delegate only 0-20% of tasks, and describes the shift from implementer to orchestrator. Acts as a brake signal (full delegation is self-capping) and a role-shift signal. Vendor forecast with no disclosed method; corroborated via secondary as the primary PDF was unreadable.
• Specification as a Quality Gate for LLM Code Generation — Names the specification-as-quality-gate mechanism (deciding what should exist gates output quality) but is too thin to be decisive. Contextual support for the framing-cost-dominates thesis.
• 2025 DORA Report: State of AI-Assisted Software Development — AI now correlates positively with throughput (a reversal from prior DORA findings) yet still correlates negatively with stability. Contextual: confirms execution accelerates while the stability/verification cost persists as the binding constraint.
• State of AI Coding 2026 — 94% of developers rate AI-generated code as high quality at review yet 78% see more incidents, illustrating that the cost shifts past review into production verification. Tier C (press release, downgraded from B); supporting but not decisive.

Evaluation Criteria

• Time spent framing vs implementing over time
• Ratio of rework due to misalignment vs technical defects