March 2025. Built with Poon Wei Kang at the AI Tinkerers Singapore TextArena Agent Hackathon. 2nd place out of ~20 teams.
the format
TextArena is a competitive environment where AI agents battle each other in text-based games. The hackathon brief: build an agent, get it to play in TextArena games, and accumulate the highest average ELO. Teams had to qualify by playing at least 10 games in each category — Public Set (announced at the start) and Private Set (revealed 30 minutes before the evaluation cutoff, giving everyone a final sprint).
Judging was split between raw ELO performance and a creativity rubric: Originality and Innovation (40pts), Strategic Thinking (30pts), Complexity and Technical Creativity (20pts), and a bonus 10pts for Use of Deviousness and Manipulation — because it's fair game as long as the TextArena environment allows it.
what we did
Most teams came in with extensive AI engineering backgrounds, sophisticated agent architectures, and elaborate MCP setups.
We went the other way.
Our approach was creative prompting and minimal MCPs — fewer moving parts, more deliberate strategy. We spent our time thinking about how to make the agent reason well about game state rather than stacking infrastructure on top. When the Private Set was revealed and everyone scrambled, the lean setup meant we could adapt fast.
We kept beating teams we had no business beating. Each round felt like we shouldn't still be in it. We were.
the result
2nd place out of roughly 20 teams.
The lesson was one we already suspected: in competitive AI contexts, prompt design and strategic clarity often outperform raw engineering complexity. A well-prompted small system beats a poorly-directed large one. The games made that visible in real time.