I’m a student at Columbia and spent the weekend building a minimal autonomous agent to play Pax Historia (the grand strategy game launched recently by YC W26).
The agent runs in a closed loop without using computer vision. Instead, I reverse-engineered the network traffic to build a perception layer.
Architecture:
- Spy: Intercepts websockets/network traffic for ground-truth game state.
- War Room: A local file system that acts as long-term memory (persisting strategy between turns).
- Brain: An LLM that ingests the state + "Constitution" (doctrine) to output plans.
- Hand: Playwright scripts that execute the move on the UI.
It managed to win the Pacific War scenario by August 1936 (20 turns) completely autonomously.
I’d love any feedback on the architecture or the interception approach! I'm still refining the loop and would appreciate critiques from more experienced engineers here!
Hi HN,
I’m a student at Columbia and spent the weekend building a minimal autonomous agent to play Pax Historia (the grand strategy game launched recently by YC W26).
Video Demo: https://x.com/PhillipYan2/status/2020912702273749050?s=20
The agent runs in a closed loop without using computer vision. Instead, I reverse-engineered the network traffic to build a perception layer.
Architecture:
- Spy: Intercepts websockets/network traffic for ground-truth game state.
- War Room: A local file system that acts as long-term memory (persisting strategy between turns).
- Brain: An LLM that ingests the state + "Constitution" (doctrine) to output plans.
- Hand: Playwright scripts that execute the move on the UI.
It managed to win the Pacific War scenario by August 1936 (20 turns) completely autonomously.
I’d love any feedback on the architecture or the interception approach! I'm still refining the loop and would appreciate critiques from more experienced engineers here!