I’m not a programmer, but I actively use AI in my areas of interest. And since my interests are heavily text-based, Antigravity turned out to be a very convenient environment for using agents somewhat beyond its intended purpose.
But imagine my disappointment when I realized I had no way to natively run a simple chain of agents that sequentially process a single document, each solving its own task. Each iteration required manual execution.
This was so strange and inexplicable to me that I began trying to find a solution for automating agents in Antigravity. No matter how much I searched or how much I called on the agents themselves for help, everything pointed to the same answer: use an API or a heavyweight framework.
I wasn’t satisfied with this situation because I needed something very basic: for the agents to pass the task down the chain without my involvement. All the prompts and skills were already ready, and I needed only one thing—not to do this manually in Antigravity.
In yet another desperate attempt to find a way to automate this, I managed to corner the agent. And we discovered a very simple, yet not at all obvious, solution.
I simply didn’t know that’s not how it’s done.
Without having written a single line of code in my life, I used agents to create an extension for Antigravity that allows agents to launch other agents. I was so impressed with the result that, forgetting everything else, I spent over two months enthusiastically developing it to share with the community.
The irony is that I’m critical of vibe-coded products—and yet I’ve become that very vibe-coder who’s bringing his creation into the world.
During this time, I learned a lot about how agents function in the Antigravity environment and developed my own approach to orchestrating reliable workflows based on this plugin, which can run for hours and involve hundreds of agents without any runtime or external control.
And all this thanks to just ~240 lines of code, with no external dependencies, two commands, and a single trigger file. But these constraints also shape the approach to building such workflows—which perhaps deserves a separate discussion.
A one-minute timelapse of a two-hour workflow demonstrating basic extension capabilities—not about OCR, but about the mechanics of Pipeline Relay using a clear example. Under the hood, there’s a simple prompt of a few dozen lines, and nothing else. An agent performs its part of the task and launches the next one.
And so on, 181 times over the course of 2 hours and 15 minutes. This isn’t a pure chain; parallel agents were launched to extract the illustrations. Each agent processes only one file or performs only one task—this is the key to stability and quality results in this architecture.
This approach is unlikely to work in every situation; it has many limitations, but, in my opinion, it also offers a number of unique capabilities. And perhaps you will find it useful too.
The duration, sequence, and completion of the orchestration are all defined through natural language prompts. But this simplicity gives rise to new complexities that require a special approach to organizing the work environment. I believe this could serve as a kind of testing ground for understanding how to manage agents.
I’ve also tried Claude Code and Codex, which have native orchestration. But for me, it’s far clearer, more convenient, and more interesting to use this particular extension. What I accomplish here with a single prompt is significantly harder to do there.
- What I like about Antigravity is that everything happens in chats; I see the full picture of what’s going on and can evaluate agent behavior in real time, rather than based on results.
- This also provides a clear structure for the sequence of events.
- There are no limits on workflow duration or the number of agents involved, except for objective constraints (tokens, connection stability, server availability).
- There is no “black box”; every agent can be an orchestrator, and I decide how the orchestration will proceed.
- No code or complex abstractions are needed; you just need to think through the logic and clearly define the prompts.
This release is a minimalist core, but there is potential for further development around this core, and perhaps the community will explore it. I already have several ideas that I plan to implement in the future, if circumstances allow.
You can install the plugin directly from Antigravity — Pipeline Relay on Open VSX
For technical details — GitHub repository
I’d love to hear your feedback.