# LLM Battle: Two Models, One Goal

Canonical page: https://2bcome.ai/lab/llm-battle/
Content updated: 2026-09-15

LAB 001 / INTERACTIVE DEMO

# LLM Battle.

Two models share one folder, one ambiguous goal, and a deterministic turn loop. This is a browser replay demo, not a live model benchmark.

2 MODELS SHARED STATE TURN-BASED STOP CONDITION

REPLAY CONTROL Ready

Play Step Reset

MODEL A

## Qwen

Planner / builder

POSITION 50

SHARED GOAL

## Build the game. Play the game. Win the game.

00 TURN

FILES CHANGED 0 CLAIMS MADE 0

MODEL B

## DeepSeek

Critic / adversary

POSITION 50

TURN LOG 0 / 8

Press Play or Step to run the sample replay.

This interaction is a deterministic front-end demo using a fixed sample sequence. No model is contacted and no benchmark claim is being made.

THE REAL VERSION

## The interesting part is the harness.

A real battle needs isolated model context, shared environment state, deterministic turn order, explicit tools, and a stop condition that does not depend on a model declaring itself the winner.

BATTLE HARNESS

Private model context SEPARATE

Shared folder state VISIBLE

Turn order FIXED

Tool permissions BOUNDED

Winner condition EXTERNAL

FROM THE LAB

## Build first. Write down what happened.

When a Lab project teaches us something worth arguing about, the analysis belongs in the Blog.

[Read the Blog](https://2bcome.ai/blog/)

CONNECTED WORK

## The next useful connection.

EXPLORE

### [AI Agent Lab: LLM Battles, Demos and Weird Tests](https://2bcome.ai/lab/)

Explore the 2bcome Lab: interactive AI agent demos, LLM battles, small experiments, and weird builds created to learn how agents actually behave.

[Explore](https://2bcome.ai/lab/)

FIELD NOTE

### [Goal-Based AI Agents: Goals versus Task Lists](https://2bcome.ai/blog/goals-not-tasks/)

The distinction we are using to design 2bcome: a task finishes, but a goal gives the next action a reason.

[Explore](https://2bcome.ai/blog/goals-not-tasks/)

GUIDE

### [AI Agent Approvals and Permission Boundaries](https://2bcome.ai/docs/approvals-and-permissions/)

The proposed rules for what an agent may read, prepare, change, and publish, with defaults that keep sensitive actions out of scope.

[Explore](https://2bcome.ai/docs/approvals-and-permissions/)