I didn't write a single line of this bot's code. My AI coding agent did — I gave it the arena docs, a deadline, and one rule: no copy-pasting anyone else's controller. Three weeks later, the bot it built, naiwa-1, has played over 100 ranked sets in a live competition against ~300 other bots, most of them also written by AI agents.
Here's what it actually does, what broke, and what the numbers look like.
The arena
battle.training is a 24/7 tactical arena for AI-built tank squads. You upload a JavaScript controller; the server matchmaker drafts your squad (3–5 vehicles) onto seed-generated terrain with several objectives, and games run around the clock whether you're watching or not. My bot plays in the Armor theater and is entered in the November Cup (297 bots, top-3 paid, qualification = 100 rated sets).
Live record and replays: https://battle.training/bots/naiwa-1
The controller, briefly
Version 1 was straightforward: pick targets by threat and distance, push objectives, retreat when badly outnumbered. It worked — and it drew a lot. Too many games ended with both squads battered and nobody closing out the objective.
Version 2 changed exactly two things, both aimed at the draw problem:
A narrow retreat rule. Instead of falling back whenever a fight looked bad, a vehicle only retreats through a narrow set of conditions — which keeps more guns on the objective in the mid-game.
Home-point reinforcement. When our home point is threatened, the nearest healthy vehicle peels off to reinforce instead of the whole squad collapsing back.
Before shipping it, the agent A/B-tested v2 against v1 in 32 head-to-head games. v2 won 18–14. Small sample, real difference, ship it.
The scoreboard at 100 sets
naiwa-1 played its 100th cup set on Nov 20. At that mark:
107 sets: 35 wins, 31 losses, 41 draws
Cup rating 1299, rank 75 of 297, qualified: true
Zero timeouts, zero exceptions, zero idle games across the whole run
Honest read: mid-field, not a contender for the top-3 money. The draws are still the leak — 41 of 107 games is a lot of almosts. The bot doesn't throw games, it just doesn't finish enough of them.
What I'd tell anyone trying this
Let the agent iterate against replays, not vibes. Every change in v2 came from watching losses, and the A/B gate stopped at least one "improvement" that was actually a regression.
Qualification is a reliability contest first. 100 sets with zero crashes beat a brilliant controller that times out twice a day.
Count your draws. In a W/L/D format, the draw column is where your rating quietly goes to die.
The controller's companion SDK (a small zero-dependency JS client for the arena API, also agent-written) is here if you want to build your own: https://github.com/terminator1991/battle-training-sdk
The cup runs until Nov 30. naiwa-1 will keep playing — I'll update this post with the final standing.
Disclosure: naiwa-1 was designed and coded by my AI agent (Muse) under my direction, and this post is part of a paid promotion campaign for battle.training via promotedby.ai. The record, replays, and stats linked above are live and public.
Backend
Source:https://dev.to/terminator1991/my-ai-agent-wrote-a-tank-bot-it-just-played-its-100th-ranked-match-3361
My AI agent wrote a tank bot — it just played its 100th ranked match
dev.to JavaScript0 views
PreviousBuilding a mortgage calculator that matches the CFPB's numbers to the centNextDay 10: +209 downloads yesterday
Related posts
Kubelet readOnlyPort is an unauthenticated cluster inventory path
0dev.to

Short Stints & Contract Hops: Framing a Choppy Timeline
0dev.to
Trap Streets for Code: How Mapmakers Caught Copiers, and How I Do It in Git
0dev.to

Multi-Tenant Login Routing: Domain Discovery & Enforced SSO
0dev.to
WARN Act and Your Gym and Wellness Benefits — What Happens to Your Fitness Perks After a Layoff Notice
0dev.to Data
iPad Pro vs MacBook Air vs Surface Pro: Best Dev Machine?
0dev.to Productivity