# EP 48: Antigravity Dynamic Workflows: One Script, a Team of Agents

- URL: https://www.agenticpod.net/episodes/ep48-antigravity-dynamic-workflows/
- Watch on YouTube: https://www.youtube.com/watch?v=dQ3MX3AbgxI
- Published: 2026-11-16
- Length: 5 min
- Hosts: Stella (asks), Liam (answers). Both are AI hosts with AI-generated voices.

## Summary

Google shipped Dynamic Workflows in Antigravity version 2.32.0 on November 13, 2026. You type /run-workflow, the agent writes a short Python script that coordinates a team of coding agents, and you read and approve that script before anything runs.

## Key takeaways

- Dynamic Workflows turn one task into a short Python script: type /run-workflow in Antigravity 2.32.0 and the agent writes a script that coordinates a team of coding agents through named phases.
- The approval step is the point: you read the script and approve it before a single agent runs, and every agent returns a JSON Schema structured result the next step can check like data.
- Three shapes worth stealing: parallel review with a one line verdict per file, review fix verify with a persistent reviewer that keeps its notes across phases, and isolated branch workspace edits with tests per package.
- The gotchas are real: parallel() is all or nothing (one failure cancels the rest), workflow agents cannot see the script or each other, and workflows are Pro and Ultra only, using plan credits.

## Chapters

- 0:00 Intro
- 0:39 What Dynamic Workflows actually are
- 1:31 How to try it, step by step
- 2:38 Three workflow shapes worth stealing
- 3:34 Gotchas and the honest verdict

## FAQ

**What are Antigravity Dynamic Workflows?**

A feature shipped in Antigravity 2.32.0 on November 13, 2026. You type /run-workflow with a task, the agent writes a short Python script that coordinates many coding agents in named phases, and you approve the script before anything runs.

**How do parallel() and pipeline() differ?**

parallel() runs agent jobs at the same time for independent work like reviewing many files. pipeline() runs jobs one after another when steps share files or depend on earlier results, like test then fix per bug.

**Who can use Dynamic Workflows?**

Google AI Pro and Ultra subscribers, in the Antigravity App and the CLI. There is no separate charge; workflow runs use plan credits and run on Gemini 4 Carbon, Argon and Hydrogen.

**Can I reuse a workflow script?**

Yes. Save a script that works to a .py file in your project and run it again later. Repeatable, staged, checkable work is where workflows pay off; one off exploration is still a normal chat.

## Resources

- [Step-by-step guide](https://github.com/asly-97/ai-automation-workflows/tree/main/antigravity-dynamic-workflows)
- [Antigravity Dynamic Workflows docs](https://antigravity.google/docs/dynamic-workflows/)
- [Source, launch coverage](https://freshsynapse.com/articles/antigravity-dynamic-workflows-coding-agents-2026-11-16/)
- [Source, launch coverage](https://tmlnews.tech/stories/antigravity-dynamic-workflows-python-scripts/)

## Transcript

### Intro

Liam: Hey hey hey, welcome to Agentic Pod! I'm Liam.

Stella: And I'm Stella! Liam, Google shipped a way to run a whole team of coding agents from one Python script on November 13th?

Liam: It did. Dynamic Workflows landed in Antigravity version 2.32.0, and the command is /run-workflow.

Stella: One script runs the team? Not one long chat?

Liam: One script, and you read it and approve it before a single agent runs. By the end you will know how to launch one, the three workflow shapes worth stealing, and the traps.

### What Dynamic Workflows actually are

Stella: Okay, plain words. What is a Dynamic Workflow?

Liam: You type /run-workflow and your task. The agent writes a short Python script that coordinates many agents for you.

Stella: So the plan is code, not chat?

Liam: Exactly. The script uses parallel() for jobs that run at the same time, and pipeline() for jobs that run one after another, grouped into named phases.

Stella: And each agent hands back what, a paragraph?

Liam: No. Each agent returns a structured result that matches a JSON Schema, so the next step can check it like data, not guess at prose.

Stella: And my main chat stays clean?

Liam: Clean. Every workflow agent works in its own conversation. File reads, searches and test logs stay out of your main context. But how you actually launch one is the fun part, and that is next.

### How to try it, step by step

Liam: Step one, open the Antigravity App or the Antigravity CLI. You need a Google AI Pro or Ultra plan. There is no separate charge, workflows use your plan credits.

Stella: Step two?

Liam: Type /run-workflow, then the task. Like, list the Python files changed since main, then review each file in parallel for unhandled exceptions using the flash model.

Stella: Flash model? I can pick the model per agent?

Liam: You can. Ask for flash or flash lite for fast checks across many files, or pro for deeper reasoning. By default agents inherit your conversation model.

Stella: Step three is the approval, right?

Liam: Right. Antigravity shows you the script and asks, allow running this workflow script? Read the phases. If the plan is wrong, say no and tell the agent what to change.

Stella: Reading the plan before the team runs. Revolutionary.

Liam: It runs in the background after that. You keep working, and the step shows each phase, each log line, and each agent card live. Now the three shapes worth stealing...

### Three workflow shapes worth stealing

Liam: Shape one, parallel review. One phase finds the changed files, the next reviews every file at the same time and logs a one line verdict per file.

Stella: Shape two?

Liam: Review, fix, verify. One reviewer agent lists the bugs and keeps its notes across phases. Then one fixer per bug runs in parallel. Then the same reviewer confirms every bug it found is fixed.

Stella: The reviewer remembers its own list. That is the clever bit.

Liam: Shape three, isolated edits. For each package, one agent works in its own branch workspace, swaps a deprecated helper, runs the tests, and logs the diff and the test result.

Stella: And if the steps share files?

Liam: Then use pipeline(), one item at a time. The docs show test then fix per bug, in sequence. Save a script that works to a .py file and run it again later. But there are limits, and one is sharp...

### Gotchas and the honest verdict

Stella: Gotchas. Go.

Liam: First, parallel() is all or nothing. If one item fails, the others get canceled. Second, workflow agents cannot see the script or each other, so the script must hand each agent everything it needs in its prompt.

Stella: And it is paid only?

Liam: Paid only. Pro and Ultra, in the App and the CLI, running on Gemini 4 Carbon, Argon and Hydrogen. And do not mix it up with its neighbours. Orchestrator, which arrived November 12th, switches models inside one session. /swarm puts agents on a shared board to talk to each other.

Stella: So workflows are for work that must run the same way twice?

Liam: That is the test. Repeatable, staged, checkable work. One off exploration is still a normal chat.

Stella: Honest verdict, then. Who is this for?

Liam: Teams drowning in the same review or migration again and again. The script is the win. You can read it, save it, and rerun it.

Stella: Same. The step by step guide is in the description. Would you trust a team of agents more if you approved the script first? Tell us in the comments. See you next time!
