Topic

Agentic Coding: How Can Systems Continuously Evolve? — AI-Native Software Engineering Practices with Lumina, Alibaba’s C++ Vector Search Library

Code is Cheap, Complexity is Not. As Agents make code generation increasingly inexpensive, the state, contracts, dependencies, and side effects of long-running projects do not disappear with the code. The codebase itself also becomes part of an Agent’s context. This talk begins by exploring how software engineering can manage complexity and what boundaries can be protected by **Static Guarantees** in Modern C++. It then examines the performance and evolution challenges that arise when vector search moves from an algorithmic problem to a systems engineering problem. Building on these topics, the talk uses **Lumina**, a high-performance C++ vector search library from Alibaba, as a case study to explain how an Agent-oriented **Harness** organizes project knowledge and engineering feedback. It concludes with an overview of Lumina’s capabilities and performance, as well as its future open-source plans. ### Agenda 1. Software Engineering in the Age of Agentic Coding 2. Modern C++ and Static Guarantees 3. Vector Search: From Algorithmic Problems to Systems Engineering 4. Lumina’s Agent Harness 5. Lumina’s Capabilities, Performance, and Open-Source Roadmap ### Key Takeaways 1. Understand why **“cheaper code” does not mean simpler engineering**: how state, contracts, and side effects in long-running projects affect every subsequent change. 2. Explore which boundaries **Static Guarantees** in Modern C++ can help protect, and which problems cannot be solved by compile-time checks or Agents alone. 3. Learn from Lumina’s experience how Alibaba’s search engine team builds a **Harness for Agents**, approaches code review in the Agentic Coding era, and addresses the limitations of **Agent Review**. 4. Understand the key trade-offs in vector search systems, along with Lumina’s core capabilities, performance, and future open-source plans.