# Gemini 4 Argon: How to Build an Autonomous StarCraft Bot

By Schnalz.net · 17 October 2026

https://schnalz.net/ai-guides/gemini-4-argon-how-to-build-an-autonomous-starcraft-bot

> Gemini 4 Argon builds an autonomous StarCraft bot by writing C++ code against BWAPI 4.4.0 in Google Antigravity CLI at high reasoning effort. The process yields a compiled Protoss bot refined through iterative practice games over a one-hour run.

## In short

- Gemini 4 Argon uses Google Antigravity CLI 1.3.5 set to high reasoning effort.
- The model develops a Protoss C++ bot against BWAPI 4.4.0 on OpenBW in one hour.
- The agent iteratively compiles code and tests against tiered practice opponents.
- Early practice against top-tier bots produces the highest tournament win rates.

## What is the Gemini 4 Argon StarCraft workflow?

Gemini 4 Argon autonomously develops a StarCraft: Brood War agent by writing and compiling C++ code inside Google's Antigravity CLI harness. The model operates within a one-hour wall-clock session to iteratively improve a Protoss bot against competitive benchmarks.

> **What you'll get**: The completed process produces a compiled C++ Protoss bot running on BWAPI 4.4.0 and OpenBW, collected directly from the bot directory when time expires.

**Gemini 4 Argon StarCraft setup at a glance**

- **CLI harness**: Antigravity CLI 1.3.5
- **Reasoning setting**: High effort
- **Bot language and API**: C++ against BWAPI 4.4.0
- **Execution platform**: OpenBW
- **Run duration**: 1 hour wall-clock time
- **Context compaction**: Triggered around 250k tokens
- **API pricing**: $2 input / $10 output per million

## How to build a StarCraft bot with Gemini 4 Argon

Building a bot requires setting up the StarSkirmish Bench environment and letting Gemini 4 Argon navigate the code-compile-test loop. The entire lifecycle runs locally through automated agent tools without human intervention during the hour.

1. Configure the CLI environment: Install Antigravity CLI 1.3.5, set the reasoning effort to high, and ensure web search and external web fetch tools are disabled.
2. Initialize the project files: Place the baseline mining skeleton bot in the working directory alongside BWAPI 4.4.0, OpenBW, three demo bot source files, the FAP combat simulator, and the BWEM map analyzer.
3. Start the agentic timer: Launch the session to start the 60-minute wall-clock countdown, which provides the agent with isolated workspace access.
4. Generate and compile C++ code: Allow Gemini 4 Argon to inspect starter code, implement combat and building routines, and call internal compiler tools to fix build errors.
5. Execute tiered practice matches: The agent autonomously runs practice games on ladder maps Heartbreak Ridge, Benzene, and Destination against tiered opponents ranging from baseline demo bots to elite community bots.
6. Analyze logs and refine logic: Let Gemini 4 Argon read match transcripts, game summaries, and bot logs to adjust tactics and keep the bot compiling.
7. Collect the final build: When the hour expires, the harness automatically extracts the bot folder without requiring a manual submit command.

## What tools and opponents does Gemini 4 Argon use?

Gemini 4 Argon receives specific developer libraries and tiered sparring partners to evaluate strategic performance. The agent utilizes local simulation libraries while competing across three official tournament maps.

**Opponent tiers and libraries available to Gemini 4 Argon**

| Category | Included items | Role in bot development |
| --- | --- | --- |
| Libraries | FAP and BWEM | Provide combat simulation and map terrain analysis |
| Tier D | Three demo bots | Offer baseline practice and starter reference code |
| Tiers C and B | Six mid-ranked human bots | Test intermediate tactical play and unit expansion |
| Tier A | BananaBrain and Locutus | Provide advanced competitive resistance |
| Tier S | Stardust and PurpleWave | Deliver elite benchmark testing on tournament maps |

## Tips for improving Gemini 4 Argon bot win rates

Benchmark traces show that the primary driver of a model's final score is its testing schedule during the hour. Gemini 4 Argon achieves better outcomes when it begins challenging the seven strongest bots early in the session.

> **Note**: Traces show that Gemini 4 Argon starts playing against the seven strongest roster bots early in the hour, even while its own bot is still losing to them.

Context window management is handled by Antigravity CLI, which condenses the conversation when usage nears 250,000 tokens so the agent can carry on.

## Limits and pitfalls of Gemini 4 Argon bot coding

Gemini 4 Argon displays significant run-to-run variance during autonomous coding sessions. According to StarSkirmish Bench data, its best and worst runs span an 18-point score gap, which is wider than OpenAI and Anthropic flagship models.

> **Watch out**: Gemini 4 Argon underperforms on advanced agentic coding compared to top competitors, and top human-written bots defeat current AI bots over 90 percent of the time.

The model is also capped at high reasoning effort rather than maximum effort. Geeky Gadgets notes that while Gemini 4 Argon integrates smoothly into workflows, specialized coding tasks remain an area where competing models hold advantages.

## Frequently asked questions

### What CLI harness does Gemini 4 Argon use for StarCraft?

Gemini 4 Argon runs inside Google Antigravity CLI version 1.3.5. The harness manages the one-hour countdown, tool access, and context compaction near 250,000 tokens.

### What reasoning effort does Gemini 4 Argon support?

Gemini 4 Argon operates at high reasoning effort in its CLI harness. This is the highest reasoning tier supported by the model, whereas competing models in the benchmark offer a max setting.

### How much does Gemini 4 Argon cost to run?

Gemini 4 Argon costs $2 per million input tokens and $10 per million output tokens through the Gemini API until December 1, 2026. After that date, the pricing is scheduled to double.

### How does Gemini 4 Argon rank on StarSkirmish Bench v0.2?

Gemini 4 Argon placed third on StarSkirmish Bench v0.2, scoring between 37 and 39 points. It functionally tied with Claude Opus 5.5 and Claude Fable 5.1, behind GPT-6 Astra and GPT-6.1 Sol.

## Sources

- [starskirmish.com: StarSkirmish Bench v0.2](https://starskirmish.com/bench/v0-2/)
- [Geeky Gadgets: Gemini 4 Argon Tested: Multimodal AI Strengths and Limits](https://www.geeky-gadgets.com/gemini-argon-3/)
- [Simon Willison: Release: llm-gemini 0.35](https://simonwillison.net/2026/Oct/13/llm-gemini/)