Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarHabr AI→ original

An AI agent built a 10-player browser shooter in 20 hours and 8.3 million tokens

A developer built the online shooter Battlecringe with an AI agent: a working engine appeared in 2 hours, and a stable version took two more evenings. In total, the agent worked for about 20 hours and generated 8.3 million tokens (chunks of code and text), and most of that time went to waiting for 10-minute automated tests. The 10-player game runs on a VPS with 1 GB of memory. The takeaway: the agent needs specifications and tests, and then it can handle even networking code.

AI-processed from Habr AI; edited by Hamidun News
An AI agent built a 10-player browser shooter in 20 hours and 8.3 million tokens
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1When idle, the server uses 0.1 vCPU and about 200 MiB of RAM, 150 MiB of which is MongoDB
Draw calls per frame after optimization39–68, down from 279–304
Key metric #3176 commits, 45 pull requests, about 28K lines, Go coverage about 91%

This developer's first attempt at an online shooter with AI ended in failure: Cursor produced raw networking code, poor ping, low FPS and almost no graphics, and the project was shelved. Now he has returned to the idea, thrown out all the old code and rebuilt Battlecringe from scratch with an AI agent. A working engine appeared after two hours of agent work, and a stable version took two more evenings.

In total, the agent worked for about 20 hours (most of that time it waited for automated tests to finish) and generated 8.3 million tokens, that is, chunks of code and text. The author published detailed statistics, and from them you can estimate how much time and resources a project like this costs.

What came out of it

Battlecringe is a low-poly browser shooter with a top-down view. It has PvP for up to 10 players, a cooperative survival mode with waves of mobs, procedurally generated maps, and "operations" with objectives such as rescuing a hostage or hunting a boss. Around the combat are the usual online-game features: clans, friends, a shop, skins, achievements and leaderboards. The server is authoritative: it decides hits, and the client only predicts player movement.

The stack is conventional. The client uses TypeScript, React, Three.js and WebGL. The server consists of two Go services (the game engine and a data API), MongoDB, Docker and Caddy, running on a VPS with 1 GB of memory and one vCPU. Deployment is automated through GitHub Actions. The author tested load with at most five simultaneous players and believes the server can handle up to 50. That is an estimate, not a load test.

How the agent got the project finished

The author attributes the success to discipline around the agent. The repository contains a docs folder with a backlog in which each feature is described separately. The working rules are written in Claude.md and have been revised several times: don't push to master, when to run tests, how to format releases.

The cycle for each task looks like this:

  • describing the task in text and the agent analyzing it;
  • if needed, a "grilling" session with questions in the chat;
  • a feature specification file and development;
  • unit and smoke tests;
  • an e2e match between scripted bots with screenshots;
  • a pull request and manual approval.

The agent was placed inside the bounds of an ordinary engineering process, and it worked within them.

The weak point is verification speed. The most complex scenarios, such as sudden death and ending a match on a timer, take about 10 minutes per run. So e2e tests were run once per pull request, which could contain several features, and the tests did not always pass on the first try.

Numbers worth noting

  • Time: two hours to a working engine, two more evenings to a stable version, about 20 hours of agent work in total.
  • Tokens: 8.3 million generated and about 945 million read, mostly from cache. The agent reads roughly 110 times more than it writes, which matters when estimating budget.
  • Code: 176 commits to master, 45 pull requests (21 of them into master), 14 epics, about 28K lines including documentation. Go test coverage is about 91%; TypeScript coverage was not measured.
  • Graphics: a 60 FPS target, with frame time at the 99th percentile of 16.8 ms. Draw calls (individual render commands per frame) on low, med and high are 39, 49 and 68, and before the fix in version 2 they reached 279–304. Triangles number 73K on low and 141K on med and high.
  • Resources: the client bundle is 284 KB gzipped, and Docker images are 23.6 MB (API) and 18.7 MB (game). When idle, the server uses 0.1 vCPU and about 200 MiB of RAM, 150 MiB of which is MongoDB.

What follows from this

First, an agent needs a process around the model: specifications, tests and strict rules in its instructions. According to the author, these are what kept the error count low. Second, the bottleneck is shifting from writing code to verifying it. If a full run takes 10 minutes, the agent sits idle, and that accounted for much of the 20 hours. Speeding up the e2e tests or running them in parallel pays off more than rewriting prompts. Third, Go services of about 20 MB and a small VPS suggest that the backend for a browser game can be cheap to run.

The caveats are obvious. This is one developer's personal case, the numbers have not been independently verified, the load was small, and part of the feature description was generated from the changelog. Still, as a benchmark for timelines, token usage and performance of a browser 3D project, the case is useful. The author promises a follow-up: a detailed walkthrough of the development cycle or of the engine's architecture.

Story

Key Questions Answered

What came out of it?

Battlecringe is a low-poly browser shooter with a top-down view. It has PvP for up to 10 players, a cooperative survival mode with waves of mobs, procedurally generated maps, and "operations" with objectives such as rescuing a hostage or hunting a boss. Around the combat are the usual online-game features: clans, friends, a shop, skins, achievements and leaderboards. The server is authoritative: it decides hits, and the client only predicts player movement. The stack is conventional.

How the agent got the project finished?

The author attributes the success to discipline around the agent. The repository contains a docs folder with a backlog in which each feature is described separately. The working rules are written in Claude.md and have been revised several times: don't push to master, when to run tests, how to format releases.

What are the key details on Numbers worth noting?

Time: two hours to a working engine, two more evenings to a stable version, about 20 hours of agent work in total. Tokens: 8.3 million generated and about 945 million read, mostly from cache. The agent reads roughly 110 times more than it writes, which matters when estimating budget. Code: 176 commits to master, 45 pull requests (21 of them into master), 14 epics, about 28K lines including documentation. Go test coverage is about 91%; TypeScript coverage was not measured.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…