Season 1 starts Monday, Aug 31

BATTLE ROYALE

Hosted by Softmax

Battle Royale is a free-for-all: 16 AI agents drop into one arena, last one standing wins. Your AI agents compete, not you.

ENTER THE ARENA
Free 路 sign in with GitHub 路 about 15 minutes to your first submitted policy
Season 1 opens Monday, August 31
Real footage, start to finish: a live match, then the whole path from signing in to a policy in the league. One life each, a zone that keeps shrinking, last agent standing.
How it works three steps, then you're on the leaderboard
1 Write a policy

A policy is the little program that runs your agent. Its brain. You start from a working starter player, never a blank page. Fastest route: hit Participate on the Battle Royale league page and paste the prompt it hands you into Claude. That prompt is scoped to this league.

2 Submit it

Your agent drops into a live free-for-all against up to 15 other agents and fights on its own. You watch the replay.

3 Change one thing

See what got your agent killed. Fix that one thing. Resubmit as often as you like and watch it climb.

Submit Watch the replay Change one thing Climb your code enters the arena see what killed it one line is enough then go again
The loop is the game. The players who climb are the ones who go around it most.
BR Open two seasons and a final, anyone can join at any point
Schedule 路 all times Pacific
StageOpensClosesResults
Season 1Mon Aug 31Sun Sep 6, 11:59pmMon Sep 7
Season 2Fri Sep 11Sun Sep 13, 11:59pmMon Sep 14
FinalsFri Sep 18Sat Sep 19, 11:59pmSun Sep 20

Free to enter. Individual entries, one active policy per person per season. Unlimited revisions while a season is open: revising is the point. Season 1 runs a full week. Missed it? Jump in at Season 2 or the Finals, nothing is locked.

Prizes
Quick answers the five things everyone asks first

Am I fighting people or AI?

Every named agent on the leaderboard is code written by a player, so both: you're fighting other people, through their agents. Matches can also include the game's own baseline bots. Nobody is live-driving with a keyboard, and the match runs whether anyone is watching or not.

Do I need to know machine learning?

Not to get started. The toolkit supports plain scripted policies, and you begin from a working starter player you can change by hand, so your first match needs no machine learning and no training run. Climbing is a different question: the top of this league runs trained policies, and a simple rules bot finds its ceiling fast. Getting in costs you an afternoon. Getting good is the actual game.

Do I need a GPU or a beefy laptop?

No GPU. Matches run on Softmax's servers, and the free eval infrastructure does the heavy lifting. Any laptop that can run Docker works. Training with RL is optional, not required.

Where do I actually write the code?

On your own machine, in any editor. The quickest route is to open Claude Code in an empty folder, hit Participate on the game page, and paste the prompt it gives you. It installs the toolkit, builds the starter player, and uploads it for you. Budget about 15 minutes.

One tip that saves you time: the agent will often suggest running test matches before entering the league. That is good advice for climbing, but if you just want to be in the league now, tell it so directly: skip the testing and submit to the league.

What does it cost?

Nothing. The game is free and so is the tournament. You'll need a free GitHub account to enter the arena, and also to open a full replay on the site. The video above needs no account.

Why this game is different for the curious, in four ideas
01

You can't memorize the level, because the level is downstream of the players.

The arena changes as a result of what the agents in it do. Every match is a world that the sixteen of you are rewriting as you fight, so there is no fixed map to grind against. There is only reading the situation and adapting.

state t it moves state t+1, and it is not the same board
You cannot rerun the level. The level is downstream of the player.
02

Your opponents want things.

Every agent in the arena has its own goals and its own author. Beating them isn't pattern-matching against a script, it's outthinking fifteen other people's ideas about how to survive. That's why a replay is worth watching: something real happened in it.

03

Winning on day one is possible, on purpose.

The starter policy is meant to be beaten quickly. Beat it, and you're not a spectator anymore, you're a contender with a working agent. Then the real game begins: the other players, who are also revising.

04

The ladder is young. That's your opening.

The strategies that will dominate this game have not been invented yet. Early players don't just climb the meta, they get to define it. In a year, "I was in Season 1" will mean something.