How Six Kuaishou Engineering Teams Turned AI Into Real‑World Business Impact

At Kuaishou, AI adoption is not confined to a few groups; engineers from modeling to platform teams are actively converting AI into concrete engineering capabilities and validating them in real‑world scenarios.

AI‑assisted coding dramatically increased output—one team reported an AI code contribution rate of 63.1%, turning days of work into minutes—but overall delivery efficiency rose only 12.8%, highlighting that personal productivity does not automatically translate to organizational speed.

Recognizing this gap, the teams redesigned collaboration around an Agent‑First platform called Agent Teammates . In this workflow, a product Agent drafts the PRD, a development Agent splits tasks, and a test Agent adds testing ideas, allowing context to flow seamlessly between stages and freeing engineers to focus on judgment and risk assessment.

Case study – Intellectual Property Pool : Traditional estimates predicted 60 person‑days for a project; using agents, the team completed it in 6 days with just one product manager, one engineer, and six agents. Agents performed the actual work in less than a day, and the team’s monthly per‑person delivery volume tripled in the second quarter.

E‑commerce “Little Assistant” project : Agents act as virtual assistants for millions of small merchants, explaining policies, auto‑enrolling merchants in promotions, and providing content recommendations. One merchant’s short‑video GMV grew from 73 ¥ to over 14 000 ¥ (66.7% share), and AI‑assisted merchants saw a daily GMV increase of more than 15% compared to a control group.

Anti‑fraud platform Monika : While a conventional rollout would require ~300 person‑days, the team built core functionality in 25 days with one engineer and a few interns. Using a specification‑driven approach (SDD + TDD), agents generated code after detailed specs, boosting risk‑detection efficiency by 50%.

Full‑stack delivery experiments : By integrating AI‑driven slicing and composition into the editing platform, a task that previously needed 20+ person‑days was reduced to 2–3 person‑days, demonstrating that AI amplifies both speed and the value of engineers’ expertise.

Push notification optimization : An Agent automatically queried data, generated SQL, and analyzed push‑duration patterns, revealing that 80.6% of local pushes lingered 10 seconds. This insight led to a PRD, parameter adjustments, and an A/B test to improve flexibility.

Harness framework for AI coding : The team defined a pipeline—demand analysis, design, test case generation, coding, self‑test, code review, deployment verification—each with clear inputs and gates. AI code generation rose from 31.67% to 84.46%, and L2+ demand involvement grew from 3.23% to 90.91%. A “dream mechanism” continuously updates knowledge, and pre‑coding test assets ensure quality, mitigating model hallucinations through verification and rollback.

Overall, the six stories show that merely speeding up code is insufficient; redesigning processes, embedding agents throughout the workflow, and enforcing quality checkpoints turn personal efficiency gains into organization‑wide improvements, enabling engineers to move from repetitive tasks to strategic problem‑solving.