Scaling AI Coding from Individuals to Teams: Xiaomi’s Engineering Practices

Background

Using AI, individual developers at Xiaomi became much faster, yet the organization did not see a corresponding speed‑up because demand translation, decision latency, and scattered context created new bottlenecks. The core question shifted from “how can an individual use AI?” to “how can an organization use AI?”.

1. Unified Workflow – VAF (Vibe Agentic Flow)

VAF provides a low‑threshold, menu‑driven workflow that guides users through a single vaf_starter markdown file, rendering a step‑by‑step menu. The process replaces ad‑hoc prompt writing with a deterministic sequence, reducing cognitive load and ensuring every step (e.g., yes to approve, e to edit) is recorded and automatically committed to Git.

“AI is not smart enough; the real challenge is getting people to use it.”

After scaling VAF to hundreds of engineers, the team observed that while code generation accelerated, the lack of shared knowledge and coordination limited overall efficiency.

2. Code Knowledge Index – VKF (Vibe Knowledge Flow)

VKF 1.0 attempted to translate code into natural‑language knowledge documents, but the double translation caused significant precision loss (missing conditions, altered call relationships). The team identified three layers of loss: code → document, document → task, and the hidden errors in document‑based reasoning.

VKF 2.0 re‑positions the knowledge base as an index rather than a full code explanation. It extracts entry points, call chains, and relevant files, then presents them as searchable metadata that AI can use to locate the exact code it needs. The pipeline consists of:

Structured scanning from entry points to infer architecture skeleton.

Expert boundary correction where domain experts validate AI‑generated partitions.

Domain aggregation that groups multiple services under a single business domain.

The knowledge is persisted as files that act as a state machine; every state transition is logged as an event, making the process auditable and replayable.

3. Collaborative Workbench – eight‑claw

Even after VAF and VKF, developers still worked in isolated AI terminals, leaving the team without a shared view of ongoing work. eight‑claw integrates AI tasks into Feishu (the company’s chat platform) using “topics” as the minimal parallel governance unit. Four entry points are provided:

Private chat – serial, lightweight tasks.

Workbench group – personal demand dashboard.

Collaboration group – parallel topics for multi‑person work.

Local dashboard – runtime visualization and task management.

Each topic maintains its own context, participants, decision boundaries, and execution state, enabling true parallelism across devices and locations. The system persists all task state as files, so a mobile device sees the same status as a desktop client.

Eight‑claw also abstracts multiple AI execution engines (Codex CLI, Claude Code, OpenCode) behind a unified ExecutionEngine interface, handling routing, failure fallback, and governance uniformly.

Design Principles & Future Directions

Lower the barrier for the majority rather than raising the ceiling for a few.

Provide AI with a code index so it can efficiently understand large codebases.

Prioritize parallelism and process visibility over isolated speed gains.

Tooling will evolve, but knowledge and collaboration are the lasting assets.

Future work includes semantic search over the knowledge base, federated knowledge across teams, and moving from task‑driven to goal‑driven AI where the system autonomously advances toward user‑defined objectives, intervening only at critical decision points.

In summary, AI accelerates coding, but the real productivity lift comes from systematic workflow unification, searchable code knowledge, and transparent collaborative execution that embed AI into the organization’s delivery pipeline.