Mohammed Ghaban Built an Arena Where AI Coding Agents Have to Prove Their Code Works
AI code can look right and still fail. Mohammed Ghaban built an arena that runs it to find out.
Mohammed Ghaban is an AI/ML developer and open-source contributor from Yemen, focused on building practical AI applications, intelligent agents, and developer tools.
His work spans Generative AI, LLMs, AI agents, machine learning, MLOps, and AI security. But one of his most interesting projects takes a very practical approach to a growing problem in software development:
What happens when AI-generated code looks correct but isn't?
That question led Mohammed to build CodeBots Arena, an AI-vs-AI programming battle arena where AI agents generate code, execute it in a controlled environment, and compete based on the results.
The project combines FastAPI, React, Monaco Editor, Elo ratings, and RestrictedPython to create an environment where AI-generated solutions can be tested rather than simply judged by how convincing they look.
From Kaggle to an AI-vs-AI Arena
The idea behind CodeBots Arena grew from Mohammed's interest in AI competitions.
“I was interested in creating a competition where AI systems could compete through programming, and I looked at the kinds of competitions available on Kaggle.”
While exploring existing competitions, he realised that he wasn't finding exactly what he had in mind.
“I realized I wasn't seeing the exact kind of AI-vs-AI coding competition I had in mind.”
That led to a simple question:
“Instead of only participating in competitions, why not try to build one?”
And that became CodeBots Arena.
Rather than simply asking an AI model to produce a solution, Mohammed wanted to create a system where AI-generated solutions could actually compete, execute, and be evaluated.
Why the Sandbox Matters
There are two particularly interesting parts of CodeBots Arena: how the agents are evaluated and how their code is executed.
The project uses an Elo rating system to track the relative performance of AI agents over repeated matches.
As Mohammed explains:
“ELO provides a simple way to track relative performance over time.”
But ranking the agents is only part of the problem.
The generated code also needs to be executed safely.
Mohammed makes an important distinction:
“AI-generated code cannot be treated as trusted code.”
That's why CodeBots Arena uses RestrictedPython together with validation and testing to provide controlled execution.
The result is more than two AI models generating code.
It becomes an environment where developers can experiment with AI-generated programming solutions and evaluate what actually happens when those solutions are executed.
When AI-Generated Code Looks Better Than It Actually Is
Building CodeBots Arena changed the way Mohammed thinks about AI-generated code.
One of the biggest lessons was how convincing generated code can be even when it is wrong.
“AI-generated code can look much more convincing than it actually is.”
A solution might appear logically sound but still fail because of an edge case, an incorrect assumption, or a subtle implementation problem.
That observation is central to Mohammed's interest in evaluation.
“Watching generated code actually run changes the question from ‘Does this answer look good?’ to ‘Does this solution actually work under the conditions of the task?’”
That distinction becomes increasingly important as developers rely more heavily on AI coding tools.
AI can produce code extremely quickly.
But producing code and proving that the code works are two very different things.
“Talk Is Cheap. Show Me the Code.”
Mohammed's CoderLegion profile uses the famous Linus Torvalds quote:
“Talk is cheap. Show me the code.”
But Mohammed believes the idea takes on a slightly different meaning in the age of AI-generated software.
AI can generate a large amount of code very quickly.
The real question is what happens when that code is tested.
As Mohammed puts it:
“AI can generate a lot of code very quickly, but generating code is not the same as proving that the code works.”
For him, showing the code increasingly means:
“Show me the tests, show me the results, show me the benchmark, and show me what happens when the system fails.”
That philosophy is at the heart of CodeBots Arena.
A Bug That Had Nothing to Do With AI
Not every challenge Mohammed encountered while building CodeBots Arena was related to artificial intelligence.
He also documented a PYTHONPATH issue where his tests worked locally but failed in CI.
The experience highlighted an issue every developer eventually encounters:
Something can work perfectly on your machine and still fail somewhere else.
Debugging the problem required looking at the differences between the local development environment and the CI environment.
The experience reinforced the importance of:
- Reproducible environments
- Environment configuration
- Automated testing
- CI
- Not assuming that local success means a project is finished
As Mohammed explained:
“A project can work perfectly on your machine and still fail when the environment changes.”
For someone building an experimental platform where AI-generated code has to survive real execution, that lesson is particularly relevant.
Beyond CodeBots Arena: Benchmarks and AI Evaluation
CodeBots Arena is only one part of Mohammed's work.
His profile highlights ongoing work in AI/ML, LLMs, AI agents, benchmarking, Kaggle projects, open-source development, and machine learning experimentation.
He has also worked on projects involving Legal AI hallucination benchmarking.
His interest in benchmarking comes from the same fundamental question behind CodeBots Arena: how can we measure whether an AI system is actually performing well?
As Mohammed explains:
“A model can sound intelligent while still hallucinating, reasoning incorrectly, or failing on specific edge cases.”
Benchmarking makes those weaknesses measurable.
Instead of simply asking whether an AI response sounds intelligent, benchmarks can help reveal where a model succeeds, where it fails, and how its performance changes.
Learning Through Curiosity
Mohammed has accumulated significant experience through hands-on learning.
His profile highlights 87+ Google Developer Program badges and achievements, more than 115 completed Google Skills learning activities, and 33,402 Google Skills points, including Diamond League status.
His learning covers areas including:
- Google Gemini
- Google Cloud
- AI agents
- Google ADK
- Machine learning
- Transformers
- Fine-tuning
- MLOps
- AI security
- Model evaluation
But Mohammed doesn't see the badges as the main objective.
“My learning pace comes mostly from curiosity rather than trying to collect badges.”
That mindset comes through in the projects he builds.
Rather than learning concepts only theoretically, he tends to turn them into experiments, benchmarks, open-source projects, and working software.
Building Without a Computer
One of the more unusual parts of Mohammed's journey came when his computer stopped working.
Because of his circumstances and the situation in Yemen, he wasn't able to immediately replace it.
Instead of stopping his work completely, he continued primarily using his iPhone, browser-based tools, and free resources.
It wasn't easy.
“It has certainly been more difficult and sometimes frustrating, especially when working on technical projects that would normally require a full development environment.”
But he decided not to allow the lack of hardware to completely stop him.
It's a reminder that developers don't always get to build with the ideal setup.
Sometimes you work with whatever you have available and keep moving forward.
What Mohammed Is Exploring Next
Mohammed's current interests include:
- AI agents
- Multi-agent systems
- LLM evaluation
- Generative AI
- AI-generated code
- Google Gemini
- Google ADK
- MLOps
- AI security
- Fine-tuning and Transformer architectures
He is particularly interested in systems where AI agents can take actions, use tools, write or execute code, and then be evaluated based on what they actually accomplish.
That makes CodeBots Arena a natural foundation for his future work.
Mohammed's vision for the project is clear:
“A year from now, I'd like CodeBots Arena to be a more mature open-source project with stronger evaluation, more interesting competitions, and more opportunities for other developers to experiment with AI coding agents.”
Build Before You Feel Ready
When asked what advice he would give other developers, Mohammed keeps it simple:
“Build before you feel ready.”
His approach is straightforward:
Start with a small problem.
Experiment.
Fail.
Debug.
Try again.
He doesn't expect every project to succeed.
“I don't know exactly where every project will lead, and not everything I build will succeed, but I want to keep creating, learning, and improving.”
That attitude is perhaps the strongest theme running through his work.
Whether he's experimenting with AI agents, benchmarking models, competing on Kaggle, contributing to open source, or building CodeBots Arena, the focus remains on turning ideas into something that can actually be tested.
Proving What AI Can Actually Do
AI-generated software is becoming increasingly capable.
But capability isn't the same as reliability.
Mohammed Ghaban's work with CodeBots Arena explores an important question for the next generation of software development:
What happens when we stop judging AI-generated code by how good it looks and start judging it by whether it actually works?
His answer is to build an environment where the code has to prove itself.
Generate it.
Run it.
Test it.
Measure it.
And learn from what fails.
That's the idea behind CodeBots Arena — and it may become an increasingly important way to think about AI-assisted software development.
