Skip to content

LucasOl1337/The-Last-Arrow

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

202 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

The Last Arrow ⚔️

A 2D combat game prototype in Unity / C# (local + online multiplayer) — and, more interestingly for engineers, a sandbox for LLM agents that play the game in real time.

I built it as a testbed for one idea: can an external AI agent perceive live game state, decide, and act through a stable interface — fast enough to actually fight? The combat game is the environment; the agent loop is the point.

Stack: Unity · C# · Python (agent broker) · local HUD/observability (localhost:8050)

The AI-agent loop (what makes this repo worth a look)

Bots aren't scripted state machines baked into the game. They run outside the engine and talk to it through a broker:

Unity (captures live fight state) → broker → agent → Codex-backed model → broker → Unity (applies the action)
  • Perception → decision → action over a clean boundary — the game streams state out; the agent streams actions back. The engine never hard-codes bot logic, so the "brain" is swappable without touching gameplay code.
  • Bot Manager — dual bot slots with persistent profiles and memory, launched from a single entrypoint (python mainbot.py).
  • Observability built in — a technical HUD, per-slot overlays, and a local docs panel so you can watch what the agent perceives and why it acted.
  • Automated playtesting — bot matches generate data that feeds back into tuning combat feel and balance, instead of relying only on manual play.

It's the same instinct behind my production work (agents that operate a system through a defined tool boundary, with logging and memory) — here applied to a real-time game instead of a SaaS.

The game itself (honest scope)

A prototype (v0.3.0), not a finished title:

  • Local and online 1v1 combat, "first to 5 kills" format with round indicators.
  • Playable characters with distinct identities (a fast samurai; an electric martial artist).
  • Modular architecture split into Core · Gameplay · Input · Match · Presentation.

Direction: sharper combat feel and visual feedback, stronger character identity, and using the automated-playtesting data to drive what changes next.


Built by Lucas Oliveira — full-stack / AI engineer. More work: LojaSync (ERP automation, React + FastAPI) · ChessCam (real-time multiplayer chess).

About

PvP arena game in Unity/C# (TowerFall-style) — my sandbox for real-time gameplay and AI playtesting, including LLM-driven bots bridged from Codex CLI into Unity.

Topics

Resources

Stars

3 stars

Watchers

0 watching

Forks

Packages

 
 
 

Contributors