AI Agent product manager focused on enterprise agents and agentic workflows. I think a product is only as good as the benchmark that defines it. Portfolio and long-form essays.https://sunxin.xin/https://sunxin.xin/en/blog/agent-demo-production-gap/https://sunxin.xin/en/blog/agent-demo-production-gap/Codex helped me build a convincing social media agent demo in very little time. Getting it ready for real users meant making the system consistently check, constrain, and take responsibility for what the agent produces.Wed, 22 Jul 2026 00:00:00 GMTAI AgentProductMethodologyhttps://sunxin.xin/en/blog/loop-engineering-rl/https://sunxin.xin/en/blog/loop-engineering-rl/The last piece was the plain-language version; this one is hard mode. My claim is simple — Loop Engineering is just reinforcement learning moved to inference time with the weights frozen, and once you accept that mapping, every design point and every way a loop crashes falls straight out of RL.Mon, 22 Jun 2026 00:00:00 GMTAI AgentReinforcement LearningMethodologyDeep Divehttps://sunxin.xin/en/blog/loop-engineering/https://sunxin.xin/en/blog/loop-engineering/I write a prompt, watch it fail, paste the error back, try again, and twenty minutes later it hits me — I'm not using AI, I'm babysitting it. Loop Engineering, this year's hottest term, is about getting out of that trap, and which jobs you should never hand off this way.Sun, 21 Jun 2026 00:00:00 GMTAI AgentMethodologyDeep Divehttps://sunxin.xin/en/blog/agent-capability-boundary/https://sunxin.xin/en/blog/agent-capability-boundary/After a few years building with large language models, I keep landing on the same conclusion. What decides how much you can hand to an AI agent isn't how smart it is. It's whether you can check its work quickly and reliably.Fri, 12 Jun 2026 00:00:00 GMTAI AgentMethodologyDeep Divehttps://sunxin.xin/en/blog/ai-native-person/https://sunxin.xin/en/blog/ai-native-person/Since 2023 I've been building things on top of large language models, and I'm more and more convinced of one thing — the people who adapt to AI fastest usually aren't the most technical. Every generation of "native" people differs not in their tools but in their default assumptions.Fri, 12 Jun 2026 00:00:00 GMTAI nativecognitionlong readhttps://sunxin.xin/en/blog/agent-ui-counterintuitive/https://sunxin.xin/en/blog/agent-ui-counterintuitive/A nine-year-old built a polished game with Claude Code. The model is already smart enough. The real question is what kind of interface lets an eager beginner steer all that intelligence.Tue, 09 Jun 2026 00:00:00 GMTProductAI AgentInteraction DesignLong Readhttps://sunxin.xin/en/blog/ai-era-pm/https://sunxin.xin/en/blog/ai-era-pm/PRD is dead, long live the benchmark. In the AI era, the thing a PM uses to define a product's value is shifting — from writing requirement docs to defining standards for how a model should behave.Tue, 02 Jun 2026 00:00:00 GMTProductAIEssay