What Is the Antigravity Swarm? Gemini 4 Carbon and Factorio
Google's Antigravity swarm let Gemini 4 Carbon reverse engineer Factorio and run agents to train-building. What it shows about speed, polish and refusals.
For Engineering, Security and Operations

What is the Antigravity swarm?
The swarm is a new feature in Google Antigravity, Google's agent app for coding and long projects, available as a desktop application and as a command line tool. Instead of one AI working through a task step by step, a swarm runs several agents on a larger job. In one public hands-on test, Gemini 4 Carbon used it to reverse engineer the game Factorio, build a harness around it, and run AI agents that reached the point of building train tracks and trains.
TL;DR
- The Antigravity swarm is a new feature in Google's agent app, usable from the desktop application or the command line.
- In a creator's test, Gemini 4 Carbon reverse engineered Factorio, built a harness for language models to play it, and ran agents through it. Many runs reached train building.
- It mostly one-shotted the build. One fix was needed because the game replay would not display, which took a second round.
- At high effort it was much slower than rivals. The creator estimated roughly three to four times as long as Opus.
- It skipped polish and scope creep. It built the working engine and stopped, and it did not design a careful benchmark around it.
- According to the creator, it refused reverse engineering requests less often than Fable and Opus, and performed on par with Astra.
- This is one person's early test, not a controlled benchmark. Treat the findings as signals to try on your own work.
What happened in the Factorio test?
Factorio is a factory-building game where you mine resources, automate production, run research and eventually build trains and a rocket. The creator gave Gemini 4 Carbon, the largest of the three Gemini 4 models mentioned (alongside Hydrogen and Argon), a hard assignment: work out how the game functions, then build a way for language models to play it.
According to the creator's account, the model:
- Reverse engineered the game well enough to understand how to control it.
- Built a harness, using some existing mods that were already out there.
- Ran agents through the simulation. They mined resources, set up industry and research, and worked toward the first goal of building train tracks and trains.
- Succeeded across multiple runs. A lot of the runs made it through.
The creator described the build as close to a one-shot. The exception was the replay: the game's replay could not be viewed inside Factorio, so the model needed a second round to correct it. Solving problems such as how to place rails was part of what it handled along the way.
The creator also began a follow-up request to see whether the agents could reach the rocket launch stage, which is essentially the end game. The excerpt does not show how that turned out, so there is no result to report.
What is a harness, and why does reverse engineering matter?
A harness is the scaffolding around a program that lets an AI use it. It tells the AI what is happening (the game state), gives it a way to act (commands), and records what happens. Without one, a language model that only produces text cannot play a game or operate software.
Reverse engineering means working backward from a finished program to understand how it behaves, without having a manual. That is the hard part here, because Factorio was not designed for AI players. Building a harness means first figuring out how to read from and write to the game.
The same pattern shows up outside games. Teams use it to understand old internal systems with no documentation, to connect software that has no official integration, and in security research. The Factorio test is a playful example of a skill with serious business uses.
Why was it so slow?
The creator ran Carbon at a high effort setting, and the difference was obvious quickly. It thinks longer and runs longer than comparable projects with Astra or Fable, and the creator estimated about three to four times longer than Opus. That is a personal estimate, not a measured benchmark.
Speed matters differently depending on your job:
- If you hand over a big task and walk away, slower thinking costs you little.
- If you are iterating live, with a back-and-forth every few minutes, it will feel sluggish.
- If you pay per token or per usage, longer runs may cost more. The source gives no pricing, so check before assuming.
How polished was the result?
Not very, and the creator considered that a feature in this case. Carbon "aggressively gets to the finish line." It built the working engine, ran the agents, produced the train and the train yard, and reported it was done. It spent little time documenting the work or thinking through how to design the harness as a proper benchmark.
The creator contrasted this with Fable and Astra, which they said tend to get carried away, with scope creep that can be extreme. Neither style is better in general:
| Style | What you get | Best for |
|---|---|---|
| Finish line first (Carbon, per the creator) | Working result, little extra | Prototypes, experiments, tight scopes |
| Thorough and expansive (Fable and Astra, per the creator) | More polish, more documentation, sometimes more than you asked for | Deliverables others will use |
If you want polish from a model that skips it, you can ask for it in a second step. If a model over-delivers, you can fence it in with a narrow instruction. Either way, say in your request how finished the result needs to be.
Does it refuse reverse engineering less often?
The creator's experience was that Fable and Opus refuse more often when asked to reverse engineer certain things, while Astra has no such reluctance. Carbon performed in the same range as Astra, and the creator could not say which was better. They described both as excellent.
Refusals matter for legitimate work. A security analyst studying malware, an engineer documenting a legacy system, or a developer building an integration may hit a model that declines. A model that refuses less is more useful for those jobs, but it also shifts responsibility to you. Reverse engineering can touch software licenses, terms of service and local law. Check your company's policy and the terms of whatever you are analyzing before you point an agent at it. The creator's intro also teased an account ban shortly after raising a swarm in another test. The excerpt does not explain it, but it is a reminder that heavy agent use can run into platform rules.
How do you try the swarm?
Per the creator, there are two ways in:
- The desktop application. Antigravity is described as Google's answer to Codex or Claude desktop. Everything runs inside the app.
- The command line. After installing, you launch it with the command agy. It asks for permission very often. The creator mentioned a "dangerously skip permissions" option as an alternative to running everything in the desktop application, and the name signals the risk. Only use it in a sandbox or on a project where a mistake costs you nothing.
If you are not a developer, start with the desktop app and give it a low-stakes task.
What does this mean for you?
If you manage technical teams, budget for slower, deeper runs on complex jobs, and decide which tasks warrant them. Set rules about what agents may analyze, especially anything involving third-party software.
If you work in operations, analytics or IT, documentation-free systems are the obvious use case. A model that can work out how a tool behaves and build a connection to it may save weeks, if your policy allows it.
If you are not technical, the takeaway is that agents are moving from answering questions to building and running things. You do not need a swarm yet. You do need to start practicing with clear task descriptions, defined finish lines and checks on the output.
For everyone: do not pick a model from a single demo. Run the same real task from your own job through two or three models. Note how long each took, how much cleanup the output needed, and whether any refused.
Could you build your own tracker for comparing AI models?
A plain log of model, task, time taken, cleanup needed and refusals is exactly the evidence this test lacked. You could keep it in a spreadsheet, or build it as a small internal app. Remy builds and deploys full-stack apps from a plain-English description, so you could describe a task-log form with a results dashboard, have your team enter runs, and own the finished tool instead of adding another subscription. Remy also gives apps access to 435 AI models, which makes side-by-side comparisons practical.
Frequently Asked Questions
What is Gemini 4 Carbon?
It is described as the largest and smartest of three Gemini 4 models in the source, alongside Hydrogen and Argon. The creator says Hydrogen is distilled from Carbon and now powers Google Search's AI mode answers.
Is the Antigravity swarm available to everyone?
The source does not say. It shows the swarm in the Antigravity desktop application and command line tool but gives no details on access, plans or pricing. Check Google's own pages.
Did Gemini 4 Carbon beat Factorio?
Not as far as the source shows. Agents reached the stage of building train tracks and trains across multiple runs. The creator started a follow-up attempt at the rocket launch stage, but the result is not in the excerpt.
Is this a reliable benchmark?
No. It is one creator's hands-on test, with a harness the creator admits was not designed carefully as a benchmark. The speed comparison with Opus is an estimate. Use it as a prompt to run your own tests.
How we source, write, and correct our articles. Editorial standards
Related articles

How to Build a Recursive Agent Harness That Writes Tools

How to Set Up AI Model Fallbacks With OpenRouter

What Is Goal Engineering? LangChain's Idea Explained
