AI Coding in Practice: Insights from ByteDance’s VP of Technology

Metrics and Efficiency

In the past year ByteDance’s AI code contribution rate grew >6×, token consumption grew 5×, and AI‑generated code merge rate >2×. TRAE team data show that >90% of code was written by AI and per‑person throughput increased 60% (1.6×). Because AI can generate code roughly 10× faster than humans, the overall efficiency gain of only 1.6× indicates that contribution rate alone is a misleading metric.

Vibe Coding vs Harness

Vibe Coding is a lightweight workflow where a prompt produces a code version that is iteratively refined. An experiment on a medium‑complexity video‑preview feature combined three mainstream coding models with three agent frameworks, running 100 trials per combination (900 runs total). All combinations achieved >80% functional correctness, but scores on UI usability, interaction feasibility, reliability, maintainability, performance, and compatibility dropped sharply, with random failures such as poor error handling, missing component reuse, and violations of engineering norms.

Adding a “Harness”—engineering‑level foundations including context management, architectural constraints, and knowledge‑base memory—raised functional correctness to ~90% and deliverability scores from ~40/100 to ~80/100, demonstrating the need for systematic engineering support.

Collaboration Across Roles

AI coding enables product, design, and operations staff to generate code directly. A product manager built a feature with Vibe Coding, but engineering review uncovered performance, extensibility, and security issues, causing delays despite runnable code. The challenge is to allow many participants to contribute valuable code while ensuring conformity to unified architecture, standards, and delivery pipelines.

Three Focus Areas

Identify metrics that reflect global delivery efficiency and productivity rather than raw AI contribution.

Governance: adopt stable, systematic practices (“Harness”) to turn Vibe Coding into production‑grade software engineering.

Collaboration: improve coordination among downstream roles to boost overall efficiency.

Prototype‑Driven Development

Traditional development follows PRD → design → spec → code. AI lowers the cost of building interactive prototypes, turning static designs into dynamic, testable artifacts. In TRAE, a “feedback” command prototype generated by AI reduced designer effort from 60% of the time spent on visual reconstruction to near‑zero.

Systematic AI Development

Systematic AI Development embeds AI throughout the workflow: AI writes specifications, validates functionality via browser automation, performs bug‑fixes, and automatically submits and releases code.

Organizational Scaling

Early AI coding relied on a few high‑skill individuals adept at prompting and context management. ByteDance is documenting and productizing practices such as prototype‑driven development and systematic AI Development. TRAE’s token consumption now reaches 5.6 trillion per day, a 50‑fold year‑over‑year increase, and the user base is expanding to non‑technical roles. TRAE Work, integrated with Volcano Engine, extends these capabilities to broader enterprise scenarios.