Antigravity lets its lead agent draw up a Python script that puts a whole crew of subagents to work at once
With version 2.32.0, Google's development environment gets Dynamic Workflows, an approach Claude Code already introduced in May.
Google updated Antigravity to version 2.32.0 on 13 November, the official changelog says. The main addition is called Dynamic Workflows: agents can coordinate parallel subagents and multi-step tasks through Python, using the Python runtime bundled with the app. Google warns that new versions roll out gradually and may take a few days to reach everyone. The same update also better protects the browser tabs an agent controls: embedded frames from other sites load without a signed-in user and stay out of the agent's screenshots unless explicitly allowed.
AlphaSignal explains how it works. With the /run-workflow command, the lead agent writes a Python script and first shows it for approval. The script then runs in the background while the conversation remains available, and the lead agent summarises the results. Each agent in the workflow gets its own conversation and cannot see the script or the other agents, so the script has to pass on all the context needed. A different model can be chosen per agent, and each agent can work in its own Git worktree. According to AlphaSignal, the feature is available to Google AI Pro and Ultra subscribers at no extra charge, although workflows with many agents can quickly burn through model credits.
The idea is not new. Anthropic's Claude Code already gained dynamic workflows at the end of May, in which Claude writes a script that directs dozens to hundreds of subagents, according to Claude Code's official overview of new features.
Transparency: this edition was compiled by Claude, made by Anthropic.