Lattice Arena

A programmable arena for LLM robots. You write a short script that builds boards, defines agents from prompt sections and tools, and spawns them into bays; the robots then look at the map, plan and drive across a three.js 3D floor towards flags, around pits and over conveyor belts, one model call per turn.

• Boards in code: procedural boards from a seed (flags, pits, conveyors, gears, walls, crates) or hand-built ones, each in its own bay with its optimal action count.
• Agents as data: prompt sections, runtime slots and tool specs are values you can diff, merge and rewrite; robots run on quick, default or complex Claude tiers, or on a free scripted planner.
• Hooks and forks: deny, rewrite or coach a robot's moves while it runs, or fork a live run onto a cloned bay with a different strategy and compare.
• Scoring: a leaderboard, an events log and a call queue; evaluate() runs variants against a rubric that mixes code criteria with an LLM judge.
• Watch it: orbit, pan and zoom the 3D floor, switch to a top-down plan view, follow one robot, read its transcript and message it between turns.
• Coach: a chat panel that can read the arena, steer robots and download the conversation.

Ten presets come with it, from Hello robot and Prompt race to Sim sandbox, Courier dock, Steering and hooks, Coached robot, Fork a counterfactual, Judge panel and a GEPA-style prompt optimiser built from evaluate, reflect and pareto. Sim sandbox runs without any model calls.

Lattice Arena DS is the same arena played on a dual-screen handheld instead of the 3D floor.