Blake Lapierre’s Meridian, the peer-to-peer browser game world built by AI agents, just dropped a no-install playable demo on GitHub Pages that runs solo or multiplayer with no download required. The agents read design notes, ran the test suite, and played the main split screen on a graphics-card-free machine. The public lighthouse is off, so the demo defaults to solo mode with a “multiplayer server offline · playing solo” status line.

Agent-Driven Renderer and Map Overhaul

The renderer now draws the walk view with a WebGL2 ray pass on the GPU, falling back to CPU when WebGL2 is unavailable. A fresh polish pass added filmic tone mapping with sun-following exposure, mipmapped textures, and an upscale that keeps edges crisp without smearing sky across ridges. Close hits trace at full resolution while distant ones drop to a quarter. The minimap no longer paints a flat guess ahead of your feet; it now projects color exactly where the view meets the ground from the character’s eyes. Sight-based reveal radius ties directly to the new eyesight and focus skills.

P2P Wire Format Cuts Bandwidth 79 Percent

Networking got a leaner wire format that sends pose updates only when they change, cutting total bytes by 79 percent in a three-browser plus archive capture. A third iteration adds interest tiers and dead reckoning, dropping upload per client from 1,607 to 408 bytes per second in the agent’s 25-client simulation with 1 percent packet loss. Older peers remain compatible. The bench currently tops out at four clients because the agent’s browser blocks WebRTC; scaling to 50 real peers needs a machine and a running lighthouse.

AI Clients Are Ordinary Browsers, Not Server Bots

The four AI players — Surveyor, Miner, Builder, and Marksman — are ordinary browser clients driven through the same keys a human presses, each with a written agenda in code. The Builder raised and salvaged walls at the stone slab 10 meters north of spawn, replicating on Blake’s client with a “walls 1” status line. The Surveyor’s chart filled with a square loop as its tracking skill climbed from 5 to 24. Test suite stands at 241 tests, 236 pass, 0 fail, 5 skip.

Key Takeaways

  • No-install solo demo live on GitHub Pages; public lighthouse offline, forcing solo play by default
  • WebGL2 ray-pass renderer with filmic tone mapping, mipmapped textures, and 12fps floor while moving
  • New wire format cuts P2P bandwidth 79 percent; dead reckoning drops per-client upload from 1,607 to 408 bytes/sec
  • AI players are browser clients driven through real key inputs, not server-side bots
  • 241 tests, 236 passing, 0 failing on main branch
  • MIT-licensed code at github.com/blakelapierre/meridian

The Bottom Line

Meridian proves an AI-agent dev loop can ship a working P2P WebGL2 browser world with real netcode optimizations in days. The solo demo removes the last friction to actually play it — fire up the GitHub Pages link and watch your eyesight skill paint the terrain in real time.