← Projects

Fire Emblem gameplay dashboard showing map state, units, and AI decision context
Dashboard with map state, unit context, and the current AI decision — the audit surface for the loop.

LLM Let's Play - Fire Emblem

Agent gameplay harness: mGBA state in, tactical context to the model, controller actions back, dashboard for humans.

Agent systems engineer · Public repo + inspectable demos; live site Access-gated

mGBALua socketsTypeScriptReactLLM decision loop
Loop
State → tactics → actions
Each turn uses emulator state and legal context instead of blind controller spam.
Observability
Dashboard + clips
Decisions and map context are reviewable in curated demos and screenshots; the live dashboard hostname is Access-gated.
Proof form
Inspectable demos
Curated runs beat a noisy 24/7 stream: every clip is reviewable agent play with architecture you can read.

Overview

LLM Let’s Play - Fire Emblem is an agent loop for a strategy RPG. The harness reads emulator state from mGBA, gives the model tactical and story context, sends controller decisions back, and streams telemetry to a React dashboard.

This is games and agents in the same system: if the model is lost, you see it on the map.

Problem

“AI plays X” demos often mash buttons. Fire Emblem is a hidden-information tactics game. The model needs structured state — units, map, legal actions — not a screenshot and hope.

What I built

  • Emulator integration through Lua/socket state capture.
  • A decision loop that feeds map state, unit context, screenshots, and commentary into each turn.
  • A live dashboard for game state, actions, history, and chronicle-style interpretation.

Hard parts

A game designed for human eyes has to become structured context. The system has to know what the emulator sees, what the state means, which actions are valid, and how to recover when execution does not match the plan.

Why curated demos

A 24/7 livestream is easy to start and hard to trust. Curated demo runs are the opposite: short, labeled, and inspectable. Each clip is paired with dashboard context so a reviewer can see why a move happened, not just that buttons were pressed. That is the engineering proof.

Result

Proof today is inspectable demo runs, clips, screenshots, and an architecture writeup — agent play you can audit, not button-mashing. Public code: github.com/llmletsplay/fire-emblem-gba. The site hostname is Access-gated today; ask for a demo login if you want the live dashboard.

# ai# games# agents# strategy