An AI agent can pass its benchmarks, perform well in demonstrations, and still fail the people it was built to help.
In this episode of CoderLegion Developer Spotlight, host Tom Smith sits down with Yaoshen Luo, a software developer and builder from Shanghai, China, to explore AI agent evaluation, real-world testing, voice AI and the challenges of building technology that genuinely works for people.
Yaoshen shares an important lesson from his experience building and evaluating AI agents: passing a benchmark doesn't necessarily mean a product is useful in the real world.
From discovering programming at age 10 to exploring AI agents
Yaoshen's interest in programming started when he was just 10 years old. A simple programming exercise in a computer class introduced him to the idea that he could write instructions and make a computer do something.
That early experience eventually led him to study software engineering at university and graduate in 2014.
After entering the software industry, he worked on mobile software and gained experience using machine learning to solve practical problems.
Today, his interests include AI agent evaluation, benchmarking and understanding why AI systems sometimes struggle outside controlled testing environments.
When an AI agent passes the test but users reject it
One of the most interesting parts of this conversation is Yaoshen's experience building a voice agent for volunteers who interview elderly people.
The volunteers needed to record conversations and complete forms containing information gathered during their interviews.
Yaoshen developed a voice agent to help with this process. The tool could record audio, convert conversations into text and help populate predefined forms.
However, there were two significant challenges:
- Speech recognition: Recognising local languages and dialects spoken by elderly people was difficult.
- Evidence and accuracy: AI-generated reports sometimes contained information without sufficient supporting evidence from the actual conversations.
But there was another unexpected problem.
Even though Yaoshen had improved the tool and demonstrated that it could perform well, volunteers refused to use it.
Why?
Because the tool introduced additional steps into their existing workflow. Instead of making their work easier, volunteers had to use the tool to generate text and then copy that text into another application on their phones.
It was a valuable lesson: a technically successful AI agent isn't necessarily a successful product.
What we discuss in this interview
- How Yaoshen discovered programming at the age of 10.
- His journey from software engineering into machine learning and AI.
- What AI agent benchmarks are and why developers need them.
- Why existing benchmarks don't always reflect real-world workflows.
- Building a voice agent to support volunteers interviewing elderly people.
- Speech recognition challenges involving local languages and dialects.
- Why AI-generated reports need evidence and reliable information.
- How real-world failures can become valuable benchmark test cases.
- Why volunteers rejected an AI tool that performed well in testing.
- Why user feedback matters when evaluating AI products.
- What Yaoshen is currently exploring in hands-on AI agent benchmarking.
- His advice to developers building their first AI agents.
An important lesson for AI builders
One of the strongest takeaways from this conversation is that benchmark scores alone cannot tell us whether an AI agent is genuinely useful.
Real users, their existing workflows, practical usability and feedback all matter.
Yaoshen also shares advice for developers starting their AI journey: use AI as a mentor and teacher, but don't stop doing the actual work yourself. Hands-on experience is essential to learning.
Watch the full interview
🎥 Watch Yaoshen Luo's Developer Spotlight interview:
Watch on YouTube
Meet the participants
Guest: Yaoshen Luo
Software developer and builder based in Shanghai, China, interested in AI agent evaluation, real-world benchmarking and practical AI applications.
CoderLegion profile:
https://coderlegion.com/user/YaoshenLuo
Host: Tom Smith
Tom Smith hosts this CoderLegion Developer Spotlight conversation, exploring Yaoshen's development journey, projects and experiences building AI systems.
CoderLegion profile:
https://coderlegion.com/user/Tom Smith
About CoderLegion Developer Spotlight
CoderLegion Developer Spotlight is a recurring interview series featuring developers, software engineers and independent builders from around the world.
We give members an opportunity to share their journeys, showcase their projects, discuss technical challenges and exchange lessons learned through real development experiences.
Are you building something interesting?
We'd love to hear your story. If you're a developer, engineer or independent builder with a project or experience worth sharing, get in touch with us about participating in a future CoderLegion Developer Spotlight.
🌐 Website: https://coderlegion.com
What do you think matters more when evaluating an AI agent: benchmark performance or real-world user experience?
Have you ever built something that worked perfectly in testing but didn't work as expected when real users tried it?
Share your thoughts in the comments. We'd love to hear your experiences.
CoderLegion #DevStories #DeveloperSpotlight #AIAgents #ArtificialIntelligence #SoftwareEngineering #MachineLearning #AIDevelopment
