Google Antigravity Dynamic Workflows: read the script first
Google Antigravity 2.32.0 adds Dynamic Workflows: the agent writes a Python script that runs parallel subagents. Read the script before you approve it.
Google Antigravity Dynamic Workflows let your coding agent write a Python script that starts many more agents. Google shipped the feature in Antigravity App 2.32.0 on November 13, per the Antigravity changelog. You type /run-workflow, the agent writes the script, and the script runs subagents in parallel and in sequence. The Dynamic Workflows docs show an approval step. They show no limit on agent count and no cost note. So the approval step is your budget control.
What actually happened
The changelog says 2.32.0 has 11 improvements and 13 fixes. The headline item: you can "orchestrate parallel subagents and multi-step tasks in Python" with the Python runtime that ships in the app. Per the docs:
- Access: Google AI Pro and Google AI Ultra plans, in the Antigravity App and the Antigravity CLI.
- Primitives:
phase()names a stage.parallel(items, fn)runs all items at the same time.pipeline(items, fn)runs them one at a time.agent()starts a subagent. - Typed results: each agent works in its own conversation and returns a result that matches a JSON Schema.
- Per-agent control:
modelacceptsinherit,flash_lite,flashorpro.workspaceacceptsbranch, which gives the agent its own Git worktree. The built-inresearchagent type is read-only. - Failure rule: in
parallel(), if one item fails, the others are cancelled. - Approval: Antigravity shows the script and asks before it runs. An approval applies to that exact script. Each subagent also obeys your permission preset.
- Reuse: you can save the script as a
.pyfile and run it again later.
Why Dynamic Workflows matter for your business
This is the same pattern Anthropic shipped in Claude Managed Agents in October. Two large vendors now agree: the unit of agent work is a script, not a chat. That is good news. A script is something you can read, review, put in Git and run again.
It also changes where the risk is. One prompt can now start 40 agents. Three things to do before your team uses it:
Read the loop, not the prompt. The cost is in parallel(items, ...). Check how long the items list is. A script that fans out across each file in a repository is a very different bill from one that fans out across five modules.
Set the model for each agent. The default is inherit, so each subagent uses the model of the parent conversation. Give search and summary work to flash_lite or flash. Keep pro for the step that writes code.
Use branch for agents that write. Parallel agents that edit the same working copy will overwrite each other. A worktree for each agent keeps the changes apart, and you merge on your terms.
Do not select "always allow" for workflow scripts. The approval is the one point where a person sees the fan-out before the tokens are spent. And plan for the cancel rule: one bad item stops the full parallel batch, so make each item small and safe to run again.
Key takeaways
- Antigravity App 2.32.0 (November 13) adds Dynamic Workflows, started with
/run-workflow, on Google AI Pro and Ultra - The agent writes a Python script with
phase,parallel,pipelineandagent; you approve it before it runs - The docs show no agent-count limit and no cost note, so read the size of each
parallellist - Subagents use the parent model by default; set cheaper tiers for search and summary steps
- Use
workspace="branch"for agents that write code, and save good scripts to Git
A fan-out script is automation, and automation needs a budget and a test. We build agent workflows with cost caps, typed outputs and a review step, on a stack you can move between vendors. See how we build agent systems you own, or run the numbers on your workflow first.
Sources: Antigravity changelog, Antigravity Dynamic Workflows docs.
- #google-antigravity
- #dynamic-workflows
- #coding-agents
- #multi-agent
- #subagents
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
TSMC employee data breach traced to a cloud provider
TSMC confirmed an employee data breach in an SEC filing. The leaked data matches records at a cloud service provider. Inventory the HR data your vendors hold.
Read itSignDrop Windows zero-day: SYSTEM access, no patch, no CVE
The SignDrop Windows zero-day uses CLFS and Defender to give a standard user SYSTEM rights. No patch or CVE exists. Keep CLFS authentication on and cut admin rights.
Read it