Wanpeng Niu
R&D Manager, Baidu Comate
Wanpeng Niu graduated from Jilin University and currently serves as an R&D Manager at Baidu Comate. He is responsible for advancing Baidu’s intelligent R&D initiatives, with a focus on driving the adoption of AI-native development paradigms across large-scale engineering systems. His work explores human-AI collaborative development models centered around Coding Agents. He holds more than 10 domestic and international invention patent applications in the field of intelligent R&D and is a core member of a key research project under China’s Ministry of Industry and Information Technology, “Intelligent Development Tools for Industrial Applications Based on Large Language Model Technology.” Prior to his current role, he was responsible for the incubation and implementation of DevOps tools at Baidu and Gitee (Open Source China), covering the development and commercialization of platforms for project management, code management, CI/CD pipelines, artifact repositories, application deployment, and operations management.
Topic
Building the Coding Agent Flywheel: Feedback Loops, Benchmarks, and Agent Engineers
As Coding Agents become increasingly integrated into real-world software development workflows, more teams are discovering that the biggest challenges are not necessarily model capabilities, but the difficulty of continuously improving Agents. Agent behavior can be unpredictable, performance can be difficult to measure, and optimization often depends heavily on a small number of experts. Drawing on practical experience from the deployment of Baidu Wenku AI Code (文心快码), this talk will explore how to build a sustainable Coding Agent flywheel through Feedback Loops, Benchmarks, and Agent Engineers. We will discuss how to engineer feedback loops that capture real-world usage signals, build scenario-based benchmarks that reflect production development environments, and continuously evaluate Agent behavior. The talk will also explore how development teams can evolve into Agent Engineers, making Agent design, evaluation, and iteration an integral part of everyday engineering practices. Outline 1. Background & Challenges Three key challenges Coding Agents face in real-world software development: uncontrollability, unevaluability, and lack of scalability Why simply having a more capable model does not solve these problems 2. Feedback Loop: Making Agent Behavior Observable How to capture real-world user feedback beyond explicit ratings Structuring and recording Agent decisions, tool calls, and user adoption signals 3. Benchmark: Evaluation Is Harder Than Generation The gap between general-purpose benchmarks and real-world software development scenarios Practical approaches to building scenario-based, multi-dimensional benchmarks 4. Agent Engineers: Bringing Humans into the Flywheel Moving beyond traditional boundaries between frontend, backend, algorithms, and platform engineering Enabling developers across roles to participate in Agent design, evaluation, and optimization 5. Making the Flywheel Run Feedback Loops provide signals Benchmarks guide optimization Agent Engineers drive continuous evolution Key Takeaways A practical, actionable framework for engineering Coding Agents Hands-on insights into designing feedback loops and evaluating Agent performance Practical approaches to evolving software engineering organizations in the Agent era
