Towards AI‑Native: How Kuaishou’s Tech Team Shifted Paradigms and Evolved Its Organization

In the context of AI deeply infiltrating software engineering, a NBER survey shows 89% of firms have deployed AI yet average productivity gains are only 0.29%, and the DORA 2025 report confirms individual efficiency rises while organizational efficiency stalls. Kuaishou’s main‑site team of more than a thousand engineers faced a similar gap: after introducing AI‑assisted coding in 2024, code‑generation rates rose from 17% to 30% but overall delivery cycles remained flat, and only about 10% of engineers changed their work patterns.

The team defined a three‑tier AI‑native paradigm:

L1 – AI assistance : AI supplies information in specific steps.

L2 – AI collaboration : AI produces initial drafts that humans refine.

L3 – AI autonomy : AI completes tasks end‑to‑end.

L2 has become the mainstream, shortening cycles by 20‑30% compared with L1. Two evolution paths were designed: a primary "L1→L2" track for large‑scale projects and a "fast lane" that jumps directly to L3 for well‑bounded, high‑impact scenarios.

Analysis of the friction revealed two core insights:

AI speed amplifies human‑human collaboration friction.

Human‑AI collaboration introduces hidden costs such as manual bridging, context alignment, verification & correction, and capability‑boundary judgment.

These frictions stem from (a) capability constraints—model limits and verification effort—and (b) structural constraints—processes originally designed for human collaboration that now impede AI flow.

To overcome them, the team rebuilt three layers:

Information Layer

Ensure AI can access accurate, reorganized knowledge, build AI‑friendly infrastructure, and establish trustworthy validation and monitoring mechanisms.

Process Layer

Transition from a human‑centric SDLC to an Agent‑driven SDLC. Separate collaboration (human alignment, decision‑freezing) from execution (Agent‑handled deterministic tasks). Use tools to convert free‑form inputs into structured specs that agents can consume.

Organization Layer

Redefine roles: split delivery (cross‑stack "feature owners" empowered by AI) from guarding (architects maintaining standards and safety). Empower designers and product owners to self‑deliver via AI, dissolving traditional boundaries.

A concrete case is the live‑gift pipeline: AI now handles direction, storyboard, generation, and review, compressing the release cycle from 20 days to under 4 days.

Talent modeling was updated to view engineer value as "base traits × capability base × core lever," emphasizing adaptability over pure coding speed.

Overall, the study concludes that AI’s true impact requires a holistic redesign of information architecture, workflow, and organizational structure; otherwise, AI’s gains are absorbed by collaboration frictions and structural losses.