The article (“Financial-grade AI coding in practice: from ‘vibe’ programming to ‘rigorous’ development”) is InfoQ China’s announcement, dated 15 April 2026, of a talk in the “Coding Agent-driven new R&D paradigm” track at QCon Beijing 2026, held 16–18 April. The speaker is an expert engineer at Ping An Technology who has led development of core business platforms, and the article reproduces his talk outline, so its content is the speaker’s own account of his team’s practice rather than independent reporting.
Its headline claim is that in the high-frequency, complex business systems where Ping An Technology’s R&D team piloted it, code written by AI has reached 60% of the code committed to the repository. The title’s contrast between “vibe” programming and “rigorous” development frames the talk as an account of making Vibe Coding-style AI generation dependable enough for financial systems.
Key Points
- The team summarises its view of how AI drives business-platform development as three pairings — “design + specification”, “knowledge + tools” and “patterns + process” — improving model and tool capability and the team’s own engineering together.
- The talk opens with Ping An’s own coding agent, 平安爱码 (Ping An Aima), described as serving more than ten thousand developers at Ping An Technology.
- The first pillar is specification-driven development: systematic AI development guidelines broken down by front-end and back-end business functions and technical tasks, which the speaker credits with supporting the 60% share of AI-written code.
- The same pillar uses AI to turn a project’s private architecture design and coding standards into CI tools such as Lint, ArchUnit and P3C rules, so that code from people and from AI alike must pass mandatory checks.
- The second pillar is knowledge engineering: a “software project knowledge bank” that stores a large system’s key knowledge in graph, vector and document databases and exposes it to the coding agent over Model Context Protocol.
- The third pillar is Context engineering by scenario, splitting core tasks into feature development, project understanding and quality assurance, each with its own standards, knowledge and tools loaded into the agent’s context.
- The fourth pillar is asynchronous human-AI collaboration: pioneer teams set aside fixed working hours — the last two hours of each afternoon and Friday afternoons — to assign asynchronous tasks to AI, using nights and weekends, through a two-stage “How-to/To-do” asynchronous development mode.
- The speaker describes the coding agent’s evolution as moving from an “intern” waiting for instructions, to an “outsourced team” that only follows documents, and next towards a “technical partner” that plans proactively.
- Open pain points named are that many projects’ architecture is not AI-friendly, that the gap between model and codebase is bridged only by a limited context window, and that building stable end-to-end agentic workflows under the financial industry’s compliance and accuracy demands still has many breakpoints.
Context
The emphasis on turning standards into enforced checks and on specification-led generation connects to Spec-driven development and Executable team standards as covered elsewhere in this wiki. All figures here, including the 60% share, are the speaker’s own and come from a pre-conference outline.