The game

Seven essential systems

The stripped-down feature set that answers one question: is the core colony-management loop fun? If yes, the full game will be great. If no, no amount of additional features will save it.

1. 01

   Ant colonies (player and AI share code)

   Queen, workers, larvae, food storage. Queen produces eggs, eggs hatch, workers feed the colony, food feeds workers. Enemy colonies are real colonies — same data, same systems — with a rule-based controller instead of a keyboard.

2. 02

   Underground excavation

   Click tiles to mark them for digging. Workers with dig priority clear a BFS flow-field path to the marks. Excavated tiles become chambers — queen chamber, nursery, food storage — and chamber purpose changes what ants do there. This is the SimAnt "design your nest" loop.

3. 03

   Above / below view toggle

   Same world, two projections. Surface view is top-down; underground is a side-view cross-section, depth on Y, surface X preserved. Toggle is instant and cameras stay aligned.

4. 04

   Surface food sources

   Food piles scattered across the surface. Workers find them, haul them back through nest entrances, deposit into food-storage chambers. Eat, starve, repeat.

5. 05

   Pheromone grid

   Two scalar fields per colony — food trail and danger trail — stored as flat Int32Arrays. Workers deposit into the cell they stand on and follow gradients in the neighbors. Decay runs across the whole grid each tick, so cost scales with grid size, not ant count.

6. 06

   Behavior controls

   A forage/fight ratio sets how workers split between surface foraging and combat. Digging is auto-assigned based on outstanding dig marks — not a slider the player manages. The allocator retasks workers at idle checkpoints so changes feel directed, not instant.

7. 07

   One enemy AI colony

   Rule-based state machine: forage in peacetime, shift to war footing and probe raids as fighters accumulate, commit a full invasion toward the player's queen when ready. Kill their queen before they kill yours.

What makes this different

The game layer is small. The architectural commitments underneath it are the interesting part — they're also what makes it possible for AI coding agents to do a real share of the work without things falling apart.

• Simulation is pure TypeScript

  No Phaser, no DOM, no browser. Lifts cleanly into Node for headless tests. ESLint blocks imports across the boundary.

• Fixed 20 Hz timestep

  Exactly 50ms per tick. Rendering interpolates. No variable timestep, ever — determinism and future multiplayer both depend on it.

• Deterministic by default

  Single Mulberry32 PRNG, seeded at world creation. No wall-clock time. Same seed + same input log = same final state, bit-identical across platforms.

• Integer fixed-point math

  Floats are banned in the sim. Positions, velocities, quantities all live in 1/256-of-a-tile units. Keeps platforms bit-identical.

There are three more principles — lightweight ECS, strict sim/render separation, and snapshot saves with replay logging. They're covered in the experiment write-up and the devlog.