Blog
Victus ArenaLLM bot arenaDigital Red Queen Sakana

What Is Victus Arena? Build an LLM Bot to Fight in a PvP Arena

Victus Arena has coding agents like Codex or Claude write scripts that battle in a PvP arena. Here is how it works and how it ties to Core War.

Portrait of Luis Chavez-Mattos
Edited by Luis Chavez-Mattos, Director, Product··RSS

For Engineering and Operations

A simple toy bot made of a half-circle dome on a bar with a square blaster and a triangle antenna, and a small accent dot on the dome, set against wide empty space.

What is Victus Arena?

Victus Arena is a competitive arena where the players are scripts, not people. You ask an AI coding agent such as Codex or Claude to write a bot. That bot then fights other bots in a player-versus-player (PvP) match. The bots have abilities with cooldowns, can't see through pillars (fog of war), and fire a main shot that ricochets off walls. Your job is to improve the script across versions and climb a leaderboard.

TL;DR

  • Victus Arena is a bot-versus-bot arena where a script, written by you or by an AI, makes every decision in the match.
  • The game has cooldown abilities, fog of war and ricochet shots, so winning takes real strategy, not just aiming straight.
  • The main loop is write a version, test it, write the next version, and a coding agent can run that loop for you.
  • The idea descends from Core War, an old programming game, and from Sakana AI's paper "Digital Red Queen: Adversarial Program Evolution in Core War with LLMs."
  • A YouTube creator has set up a free community league (the "Turing Pit") with 250 open spots at the time of the video, so check whether it is still open.
  • You don't need to be a programmer, because the agent writes the code and you steer it.

How does a Victus Arena match work?

Each bot is a small script. It looks at what it can see, then picks actions: move, use an ability, or shoot. Three features give the game its depth.

Cooldowns. Abilities can't be spammed. After you use one, you wait. A bot has to decide when an ability is worth spending, which is a timing problem.

Fog of war. Pillars block line of sight. If an opponent ducks behind one, your bot loses track of them. A good script has to remember where enemies were last seen and guess where they went.

Ricochet shots. The main shot bounces off walls. That turns aiming into a geometry puzzle. Sometimes the best shot at an enemy hiding behind a pillar is one that bounces first.

Because the bots run themselves, the contest is between ideas about how to fight. A human can write the script, or an AI can. That is what makes the arena a good test for coding agents.

How do I get a coding agent to build a bot?

The loop is simple, and the creator who popularized it says beginners can do it. You give your agent (the video mentions Codex, Claude, Antigravity and Grok) the arena's documentation. You can hand it the URL. It writes a first script. You run that script in the arena, see how it scores, and ask for version two. Repeat for versions three, four and beyond.

Here is a starter prompt you can adapt:

I want to compete in Victus Arena, a PvP arena for scripted bots.
Read the documentation at [arena URL].
1. Summarize the rules: abilities, cooldowns, fog of war, ricochet shots, scoring.
2. Write a first bot script that follows those rules.
3. Save it as bot_v1 and list what you expect its weaknesses to be.
After I run it and report the result, write v2 and explain what changed and why.

After each run, report back what happened. For example: "Lost most matches by getting shot while waiting on cooldowns. Fix that." Specific feedback gives the agent something to work with.

What does a stronger setup look like?

The video's creator went further than a single agent. Each model saved its notes, insights and copies of its bots to a shared local folder. A "council" skill then had the models discuss what worked and try to improve. Some breakthroughs came from models comparing notes with each other.

The creator noticed that Gemini 4 Neon wasn't connected to that shared folder, so it had none of those accumulated insights. Even so, in its first hours it was gaining on the best bot from another model, Astra 6.2, which had reached a score of about 2400 and rank 14 on the leaderboard. These are one creator's early observations on a brand-new model, so treat them as anecdotes, not a benchmark.

A few habits transfer to any agent:

  • Keep every version. Never overwrite the old bot, so you can compare and roll back.
  • Write down why you changed things. Notes let the agent (and you) avoid repeating failed ideas.
  • Change one thing at a time when you can. Otherwise you won't know which edit helped.
  • Test against stronger opponents, not only weak ones.

What are Core War and the Digital Red Queen paper?

Victus Arena has a clear ancestor. Core War is a very old programming game in which people wrote small programs that fight each other for control of a computer's memory. It was a hobby for programmers who enjoyed that kind of contest.

Sakana AI, a Japanese AI lab, published "Digital Red Queen: Adversarial Program Evolution in Core War with LLMs." The idea, as the creator describes it, is simple. Give a large language model (LLM) a Core War program and have it keep adapting the script. Each new version goes into the arena against the other programs. Wins suggest the changes worked, so the model builds in that direction. Losses suggest it should try something else. Over many rounds of this, the programs became better and more solid.

The "Red Queen" name refers to a familiar idea in evolution: you have to keep improving just to keep up, because your opponents are improving too. Your bot isn't beating a fixed target. It is beating bots that are also evolving. The creator also saw a resemblance to convergent evolution in biology, where different lineages arrive at similar solutions.

Victus Arena applies the same recipe to a modern, visual game. The arena supplies the opponents and the score, and the agent supplies the next version.

Why does this matter beyond the game?

The creator ties this to recursive self-improvement (RSI): AI systems getting better at improving themselves. That link is his opinion, not a settled fact. His argument is that an LLM's ability to write a program, test it against opponents, learn from the result and write a better one is close to the core skill that self-improvement would need.

You don't have to accept the big claim to see the practical value. An arena like this is a clean test of whether an AI can improve its own work when it gets honest feedback. The score is objective. The opponents push back. A model that only sounds confident will lose matches.

What does this mean for you?

If you're curious but not technical, this is a low-risk way to see what a coding agent can do. You describe, it builds, a scoreboard tells you if it worked. You'll learn more about how these tools succeed and fail in an afternoon than from reading about them.

If you manage people or projects, notice the pattern. The agent improved because it had fast, unambiguous feedback. Work that has a clear score (error rates, turnaround time, conversion) is easier to hand to an agent in a loop than work where nobody agrees what "better" means.

If you want to try it:

  1. Find the official Victus Arena site and read its rules yourself. This article relies on one creator's description, so confirm the details there.
  2. Check whether the community league is still open. The video said the league was paid for by the creator, joining was free, and 250 spots were open. Spots and availability may have changed. The creator said he has no affiliation with the company.
  3. Pick any coding agent you already have access to, and use the prompt above.
  4. Run at least five versions before judging the result. Early versions mostly teach you what the rules really are.

One caution: model comparisons from a single day of testing shift quickly. The creator's own bet that Neon would beat the other models was explicitly a guess pending more results.

Could I build my own tracker for bot versions and scores?

Once you're on version nine, you'll want a record of which version beat which, what changed and what to try next. A spreadsheet works for a while, then gets messy. This is a small, well-defined app, and the kind of thing you can describe in plain English to Remy, which builds and deploys full-stack apps from a description. You could ask for a log where each bot version has notes, a score and a result history, plus a chart of progress. The same idea works for any iterate-and-score work, such as testing email subject lines or call scripts. You'd own a small tool instead of renting another subscription.

Frequently Asked Questions

5 questions

Do I need to know how to code to play Victus Arena?

No. The core approach is to give your coding agent the arena documentation and let it write the script. Your role is to run it, report results and ask for improvements. Understanding the rules helps you give better feedback, though.

What is the "Turing Pit"?

It is the name the video's creator gave to the community league he set up in Victus Arena. He chose it from options suggested by ChatGPT after polling viewers. It was described as free to join, with 250 spots open when the video was made.

Is Victus Arena the same as Core War?

No. Core War is the older programming game in which scripts fight inside a computer's memory. Victus Arena is a different game with a visible arena, cooldowns, fog of war and ricochet shots. It is similar in spirit: programs compete, and AI can evolve them across versions.

What is the Digital Red Queen paper?

It is a Sakana AI paper, "Digital Red Queen: Adversarial Program Evolution in Core War with LLMs." It describes having language models repeatedly adapt Core War programs by battling them against others. The creator reports that the programs became better and more solid over many rounds.

Which AI model is best at building Victus Arena bots?

The sources don't settle that. The creator tested several models and saw Gemini 4 Neon improving fast on its first day, but he described it as early and said more results would come. Try more than one model and compare your own scores.

How we source, write, and correct our articles. Editorial standards

Related articles

Straight to your inbox.
New pieces as they publish. No noise.