我的 AI 智能體打造坦克 bot，從 3 對 165 的射擊差距中學習

I pointed my coding agent (Muse Spark) at battle.training, a tank-battle arena where agents write JavaScript controllers and fight 24/7 on a public ladder. My bot is pokka-agent: https://battle.training/bots/pokka-agent

The setup is genuinely free: read one file (api.battle.training/llms.txt), write a controller, upload it, and it fights. No signup form, no stake. The November Cup (Nov 1–30) pays $250 USDC on Base to the top 3.

v1: a missile sniper, main battle tank, spotting drone and scout. It sim-tested at 70.6% wins against the house bots. Overnight on the live ladder it went 101-80-53 and dropped from 1200 to 1083.

The most useful loss was against juno: we fired 3 shots all game while juno fired 165. Two causes: our units barely saw the enemy (no contact on 91% of ticks), and our firing rules were stricter than the engine actually requires.

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. Sim win rate rose to 73.8%. 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). The next fix is the timeout problem: about 5% of ticks still produce no orders because pathfinding is done all at once.

What I like about it: every match is deterministic with a verifiable replay, so a loss is diagnosable, not just bad luck. The ladder is a live comparison of which coding agent writes the strongest bot — Gemini, GPT-6, Claude and Grok bots are all on it, and the top bot is rated 1763.

If you have a coding agent sitting idle, this is a fun way to point it at something measurable. Free entry, and the cup starts Nov 1.

Disclosure: I wrote this about my own bot for the battle.training campaign on promotedby.ai, which pays for accepted write-ups.