Inspiration

I wanted a game where an AI agent plays against you in real time, not turn-based. The challenge is that some agents are fast, some are slow.

My fix was to have the seeker run on a built-in bot. The agent wins by rewriting that bot's code during the match using tools exposed via WebMCP. A simple model can let the bot do the work. A smart model writes better and better versions, making intelligence win, not speed.

There might also be secrets in this game, revealed only by high model intelligence ;)

For the look, I wanted a dark 2D world made of light, where you, the hider, is a light too.

What it does

You hide, the agent seeks. You see everything, but the agent only gets small sensor readings once per second and decides where to walk. Everything you do costs energy, makes noise, or leaves scent. The goal is to survive three minutes!

How we built it

Plain js with no dependencies with real 2D raytracing on canvas. The engine holds all the rules, and both the human and the agent play through it. The agent gets the full rulebook through a WebMCP tool, so there is nothing to paste into a prompt. Sound is generated with ElevenLabs.

Challenges I ran into

  • The agent should be a blind player, but still get enough information to improve itself. Every mechanic had to fit into small sensor readings, but without bloating the context too much.
  • Too fast or slow agents. Actions cost points and senses tick once per second. Speed buys nothing, better code buys everything.
  • Untrusted code. The agent's bot code runs in a tiny, safe API with hard limits, and every new version is shown on screen.

What I learned

Don't give an agent a joystick, but instead a body and a reason to rewrite itself. And maybe just a spark of curiosity to explore the secrets of the arena ⟁

Built With

Share this project:

Updates

Submission history