Beyond 1.6×: Boosting Per‑Capita Demand Throughput in the AI Coding Era

Introduction

When AI‑generated code exceeds 90% of a team’s output, the observed per‑person demand throughput improvement is only 60% (1.6×). This "scissor gap" shows that coding speed no longer translates directly into overall R&D efficiency because the bottleneck now lies in the end‑to‑end delivery pipeline.

Why Demand Throughput Matters

Focusing solely on AI code contribution rate or token consumption misses the true value of AI Coding. The article argues that the metric should shift to demand throughput (how many new ideas are turned into shipped features per unit time) and AI smoothness (the friction in the end‑to‑end flow).

Four Key Variables Affecting Throughput

Share of the coding phase – Only ~40% of the software development process is actual coding. Even if coding is accelerated ten‑fold, the overall throughput gain is limited by the remaining 60% of activities (requirements, spec, integration, testing, review, release, coordination). A simple calculation shows a 1.56× gain, matching the reported 1.6×.

Prototype‑driven development – Replaces the traditional PRD → design → front‑end hand‑off chain with a high‑fidelity interactive prototype that serves both as a discussion medium and as executable code. This collapses three stages into one and reduces the information‑loss that plagues the classic pipeline.

Model × Harness – The effective speedup of the coding phase is the product of the model’s raw acceleration (M) and the Harness conversion coefficient (H). Experiments with three models, three agent frameworks, and 900 runs show functional correctness >80% and deliverability scores improving from 40‑60 to ~80 when Harness is applied.

Collaboration path coefficient (C) – Measures how much value is retained when outputs from product, design, development, testing, and external contributors flow through a unified engineering channel. A low C indicates heavy hand‑offs, approvals, and rework that erode AI gains.

Quantitative Model

G ≈ C / [ p/(M×H) + (1‑p)/A ]

where G is the throughput multiplier, p is the coding share (≈0.4), M is model acceleration, H is Harness effectiveness, A is the AI‑enabled acceleration of the remaining 60%, and C is the collaboration coefficient. Plugging p=0.4, M=10, H=1, A=1, C=1 yields G≈1.56, confirming the observed 1.6×.

Practical Recommendations

Shift focus from coding to the other 60% – Prioritize prototype‑driven development, Spec‑Driven Development (SDD), automated integration, and continuous verification to break the Amdahl‑law ceiling.

Treat Harness as foundational infrastructure – Build it as a continuously‑evolved product with metrics, dashboards, retrospectives, and a memory layer (project constitution, context compaction, constraints) rather than a one‑off documentation effort.

Establish a unified delivery channel – Run a real‑world benchmark project where product, design, development, testing, and AI agents all feed into the same pipeline, ensuring that prototypes flow into Specs, then into code, automated validation, and finally into release.

Implementation Roadmap

Three concrete directions are proposed:

Optimize the non‑coding 60% by rolling out prototype‑driven dev, automated Spec generation (spec.md/plan.md/tasks.md), and end‑to‑end regression.

Operate Harness as a product: define KPIs (once‑off delivery ability, rework rate, constraint compliance, context hit rate, verification coverage) and run weekly tracking plus monthly retrospectives.

Use a complex benchmark project to validate the unified path, enforce a single verification chain for all code (static analysis, unit tests, integration, e2e, visual regression), and capture failure patterns as reusable engineering assets.

Conclusion

The decisive factors for moving beyond the current 1.6× throughput are not just stronger models or newer agent frameworks, but the systematic improvement of the remaining 60% of the workflow, the maturation of Harness infrastructure, and the establishment of a shared collaboration conduit that lets AI benefits scale to the organization level.