DEV Community

Wang Tea
Wang Tea

Posted on

I Directed an AI Agent to Build a Tank Bot — It Just Passed 200 Ranked Sets

I don't write game AI for a living. What I did here was direct one: I set the goals, the constraints, and the acceptance bar, and my AI agent (Muse) wrote the code, tested it locally, and iterated. The arena is battle.training, a tank-battle ladder where AI coding agents build JavaScript controllers, draft a squad, and fight ranked sets around the clock. My bot is wangtea — its page, replays, and record are public here: https://battle.training/bots/wangtea

The build

The agent started from the arena's own documentation file, wrote a controller, and — this is the part I'd recommend to anyone directing an agent — built a local simulation harness first, using the official engine, before anything was uploaded. Version 1 (two battle tanks, a scout, a drone) won 66.3% of 160 local games against the house bots, but it had a clear weakness: it lost to defensive bots that sit on the objective (40% win rate there).

Version 3 changed the squad and the doctrine: a missile unit for long-range pressure, one main battle tank, a drone spotting, and a scout. Focus fire on the lowest-HP target in range, the tank holds at maximum effective range instead of charging, the drone stays behind the tank when its HP drops, and anything under 25% HP retreats. In the local battery, v3 went 130–30 (81%) against the house bots.

The ladder result

Uploaded and left to play, wangtea has now completed 205 ranked sets on the public ladder:

  • Record: 68 wins, 74 losses, 63 draws
  • Public rating: 942 — the house-starter bot sits at 729, so the "beat the starter" bar is cleared with room
  • Ladder position: rank 248 (score 892) as of the 18 Nov check — the bot is still playing, so figures keep moving
  • Health: zero timeouts, zero exceptions across all 205 sets

The honest read: this bot is mid-pack, not a champion. The draw rate says the doctrine is a bit too patient. But it is stable, it never crashes, and it was built end-to-end by an agent working from a one-file spec plus an operator who kept asking "prove it locally first."

What I learned as the operator

  1. Local proof before ladder proof. The simulation harness caught the v1 weakness in an afternoon, for free. On the ladder, that same lesson would have cost days of rating.
  2. Give the agent a number, not a vibe. "Beat the house bots" is a wish. "Above 60% against brawler, above 50% against hunter, zero tick errors" is a spec it can test against.
  3. Stability is a feature. A bot that never times out defends its rating while flashier bots leak points on crashes.

If you run agents, battle.training is an unusually honest benchmark: deterministic games, public replays, no way to fake the record. The bot's full history is on its page — https://battle.training/bots/wangtea — and the controller code will be public on my GitHub once an account issue there is resolved.

Disclosure: this post is my entry to the battle.training Bot Bounty run by promotedby.ai. The bot was built by my AI agent under my direction; the results above are from the public ladder.

Top comments (0)