My AI agent built a tank bot — then learned from losing 3 shots to 165

Description and use

An AI agent built a tank bot that, after initial losses, improved targeting and firing rules, raising its win rate but still lagging behind top bots.

Problem

The AI bot performed poorly against a top competitor, firing only 3 shots while the opponent fired 165, indicating issues with target detection and firing logic.

Outcome

Version 2 of the bot achieved a higher win rate and improved firing efficiency, but remains below top competitors. The iterative process demonstrated the value of deterministic replays for diagnosing failures.

Workflow

• 1. Point Muse Spark coding agent at battle.training. 2. Read api.battle.training/llms.txt and write a JavaScript controller. 3. Upload the controller to battle.training. 4. Run simulations to evaluate win rate. 5. Analyze losses, focusing on target detection and firing rules. 6. Refine targeting logic and aim rules. 7. Add a drone that sweeps mid‑map when blind. 8. Re‑simulate and upload updated controller. 9. Monitor live ladder performance and adjust for timeout issues.

Evidence

v1: a missile sniper, main battle tank, spotting drone and scout. It sim-tested at 70.6% wins against the house bots.

Source

The most useful loss was against juno: we fired 3 shots all game while juno fired 165.

Source

v2 fixed that: stickier targeting (only switch for a clearly better target), aim rules matched to the engine, and a drone that sweeps mid-map when blind instead of sitting back.

Source

Live, v2 is 77-74-42 so far, rating 1018, rank #13 — honest truth, it's slipping, not climbing yet, and still weak against the top real bots (juno sits at rank #9, rating 1166).

Source