Inter-office memo
- To
- Everyone using Google Antigravity
- From
- Mo, Chief Executive
- Date
- November 15, 2026, 3:10 AM CST
- Re
- Antigravity’s Dynamic Workflows
Antigravity adds Dynamic Workflows. Who gets /run-workflow, and when to use it
On November 13, 2026, Google began rolling out Dynamic Workflows in Antigravity: type /run-workflow and a task, and the agent writes a short Python script that runs many agents, in parallel or in turn, which you approve first and can save to run again. Google lists it for Google AI Pro and Ultra, in its app and CLI, not for free accounts. Claude Code has had a feature of the same name, in JavaScript, since May 28.
What Google released
Google’s Antigravity changelog lists version 2.32.0 of the Antigravity app on November 13, 2026, and its first line is the news: “You can now use Dynamic Workflows to orchestrate parallel subagents and multi-step tasks in Python using the Python runtime bundled with the app.” The same day, Antigravity’s account on X said Google was “rolling out” the feature, for “deterministic control flow instead of prompt drift.” The changelog adds that new versions “roll out gradually and may take a few days to reach all users.”
What happens when you use it, by Google’s Dynamic Workflows docs:
- The agent reads a built-in
run-workflowskill and writes a Python script for your task. - Antigravity shows you the script and asks you to approve it, unless your permission preset already allows it.
- The script runs in the background, so you can keep working in the same conversation, and a workflow step shows “each phase, log line, and agent as the script progresses.”
- When it finishes, the main agent reads the workflow’s logs and output and summarizes the results for you.
Each agent the script starts “works in its own conversation and returns a structured result that matches a JSON Schema,” so “File reads, searches, and test logs stay out of your main conversation’s context.” The agents “don’t see the script or each other,” so the script gives each one everything it needs in its prompt.
Varun Mohan, who is building Antigravity at Google DeepMind, called it “another extremely used feature internally,” and added: “You don’t even need to /run-workflow, the model should be able to decide similar to swarms.”
Two things it isn’t. It isn’t Antigravity’s old “workflows,” Markdown files of “repeatable prompt sequences” from the Antigravity IDE, which Google now asks people to migrate to skills. And it isn’t brand new to developers: Antigravity’s Python SDK added it in beta in version 0.1.22, on October 12.
Who gets it
Google’s slash-command reference gives /run-workflow a plan tier of “Google AI Pro and Ultra.” Free and Google AI Plus accounts don’t get it. The reference doesn’t say whether a Pro trial counts, and doesn’t mention Enterprise. The docs list it for the Antigravity app and CLI; they don’t list the IDE extensions or other editors.
That makes it one of six ways Antigravity now splits work across agents, each with its own plan tier, as Google’s docs and models page set them out:
| Option | How you start it | What Google says it’s for | Plans |
|---|---|---|---|
| Orchestrator | Choose it in the model selector | One conversation where Carbon or Argon plans and Gemini 4 Hydrogen does the edits and tests | Ultra; Pro gets the Argon version |
| Subagents | Automatic, or an @ mention |
“Delegating a focused side task” | All plans |
/run-workflow |
/run-workflow <task> |
The same check across many files, staged steps, or a reviewer-and-fixer loop | Google AI Pro and Ultra |
/swarm |
/swarm <task> |
“Broad tasks that split cleanly,” with workers coordinating on a shared board | Google AI Pro and Ultra |
/boost |
/boost <task> |
“Hard bugs, race conditions, algorithms, and tricky refactors” | “Paid plans” |
/teamwork |
/teamwork <goal>, then a scoping interview |
“Multi-milestone projects, repository-scale migrations” | “Paid plans” |
Google’s rule of thumb: reach for a workflow when you want “predictable, scripted control flow,” and for a swarm “If your agents need to coordinate with each other while they work, for example to agree on interfaces or review each other’s diffs.”
How to run one
- Update Antigravity. The app gained Dynamic Workflows in version 2.32.0, so update if you’re on an older one, and give the rollout a few days if
/run-workflowdoesn’t appear yet. - Type the command and your task. In any conversation, type
/run-work, press Tab, and describe the job. Google’s example:/run-workflow Find every TypeScript file changed on this branch, review each one in parallel for missing error handling, and give me a one-line verdict per file. - Read the script, then approve it. The Allow running this workflow script? prompt lets you allow it this time, for the rest of the conversation or always, or say No and tell the agent what to change. An approval “covers that exact script.”
- Watch it work. The workflow is one step in the conversation, titled with its current phase and progress, such as
workflow · Review · 3/4 agents, and it’s listed under Background Tasks in the side panel. Click an agent’s card to read its prompt, watch its tool calls or answer its requests. - Read the summary. To stop early, stop the conversation, which also stops every agent the workflow started.
You can steer the script from your prompt. Each agent runs on your conversation’s model unless you ask for a lighter tier, flash_lite or flash, for “fast, high-volume checks across many files,” or pro for deeper reasoning. An agent can be a copy of the current one, the built-in read-only research agent, or one of your custom subagents. When agents edit code at the same time, ask for branch workspaces, which give each agent “its own isolated Git worktree.”
A script that works is worth keeping: ask the agent to save it as a .py file in your project, and next time run something like /run-workflow Run scripts/review-changes.py. This is the shape of one, from Google’s docs:
phase("Discover")
files = await agent(
"List the Python files changed since main. Paths only.",
schema={"type": "array", "items": {"type": "string"}},
)
phase("Review")
reviews = await parallel(
files,
lambda f: agent(
f"Review {f} for unhandled exceptions. Reply with a one-line verdict.",
schema={"type": "string"},
role=f"Review {f}",
model="flash",
),
)
for r in reviews:
log(r)
Watch your quota and your approvals
Google hasn’t said how much quota a workflow uses. Its plans page says rate limits “correlate with the amount of work done by the agent,” and a workflow is many agents working at once. When it introduced /swarm, Google warned that “parallel multi-agent systems consume quota much faster than a standard chat.” So ask for flash agents wherever a lighter model will do. For everyday work in one conversation, Google’s quota-saver is Orchestrator, which on Terminal-Bench 4.1 reached “near-Carbon task success rates while, on average, consuming 24% less quota and completing 19% faster.”
Three more things to know before a big run:
- One failure cancels the batch. In
parallel(), “If one item fails, the others are canceled.” - Agents ask in their own conversations. Each agent follows your permission preset and “waits until you answer,” so a workflow can sit still behind a request you haven’t seen. Check the agent cards when a phase stalls.
- Turbo skips the script prompt. If your preset already allows workflows, as Turbo does, the script runs without asking you first.
How it compares with Claude Code and Codex
Claude Code introduced dynamic workflows in version 2.1.154, published on npm on May 28, 2026. Its changelog described them as orchestrating “work across tens to hundreds of agents in the background.” Antigravity’s scripts use building blocks with the same names, phase, log, agent, parallel and pipeline, in Python instead of JavaScript. By the two companies’ docs:
| Antigravity | Claude Code | |
|---|---|---|
| Script language | Python, on a runtime bundled with the app | JavaScript |
| How you start one | /run-workflow, or ask in your own words |
Ask in your own words or include ultracode; /effort ultracode lets Claude decide for every task |
| Plans | Google AI Pro and Ultra | All paid plans, the API, Bedrock, Google Cloud and Foundry; Pro turns them on in /config |
| Watch a run | The workflow step, and Background Tasks | /workflows |
| Save and rerun | Ask the agent to save a .py file |
Press s in /workflows, and it becomes a / command |
| Published limits | None in its docs | 16 agents at once by default, 1,000 per run |
| Ready-made workflows | None listed | /deep-research |
Anthropic’s docs carry the same warning about usage: with ultracode on, “each request uses more tokens and takes longer,” and on a subscription “a session with ultracode on reaches a session or weekly limit sooner.”
Codex works differently. OpenAI’s docs describe “subagent workflows,” in which Codex runs them “by spawning specialized agents in parallel and then collecting their results in one response,” but the agent decides what to spawn as it works; the docs describe no script to review, save and rerun.
What Google hasn’t said
- How much quota a workflow uses, or whether there’s a cap on agents per run.
- Whether more plans will get it. Free, Google AI Plus, the Pro trial and Enterprise aren’t listed.
- Whether it works outside Google’s own app and CLI, in the IDE extensions or in editors that use Antigravity’s server for other apps.
- Which CLI version has it. The docs list the CLI, but its changelog, through version 1.5.8 on November 13, doesn’t mention workflows.
Dynamic Workflows in Enjoy
Enjoy runs Antigravity on your Google account through the server Google offers to apps such as Zed and JetBrains IDEs, beside Claude Code, Codex and the other labs’ agents. Google lists /run-workflow for its own app and CLI, so run Antigravity’s workflows there. Claude Code’s do run in Enjoy: choose Ultracode among Claude Code’s effort levels, shown as Ultra, and Claude plans larger tasks as workflows on the Claude plan you already have. If a big job uses up one plan’s limit, you can switch agents in the middle of the conversation and carry on with another. Enjoy’s own workflows are a different thing: saved prompts that run on a schedule.
8.Questions.
What are Dynamic Workflows in Antigravity?
- A way to have Antigravity’s agent write and run a short Python script that starts many agents for you, in parallel or one after another, and collects a structured result from each. Google added them to Antigravity’s app in version 2.32.0 on November 13, 2026, and its docs list the CLI too.
How do I start a Dynamic Workflow?
- Type /run-workflow followed by your task, such as “review each changed file in parallel for missing error handling,” or ask for a workflow in your own words. Antigravity shows you the script and asks you to approve it, unless your permission preset already allows it, then runs it in the background.
Can I use Dynamic Workflows on a free Antigravity account?
- No. Google’s slash-command reference lists /run-workflow for Google AI Pro and Ultra, the same as /swarm. Free and Google AI Plus accounts keep subagents, which are on every plan, and Google hasn’t said whether the Pro trial or Enterprise includes workflows.
What’s the difference between /run-workflow and /swarm?
- Who holds the plan. In a workflow, a script decides what each agent does, and the agents don’t see each other: each reports back to the script. In a swarm, named workers and a verifier split the task and coordinate on a shared message board. Google suggests a swarm when agents need to agree on interfaces or review each other’s changes.
Do Dynamic Workflows use more of my quota?
- Google hasn’t said how much. Every agent a workflow starts is a full agent that runs on your conversation’s model unless you ask for another, and Antigravity’s quota tracks the work its agents do. For high-volume checks, ask for the flash or flash_lite tier in your prompt.
Is it the same as Claude Code’s dynamic workflows?
- It’s the same idea under the same name, with building blocks of the same names, such as agent, parallel and pipeline. Claude Code added dynamic workflows on May 28, 2026, writes them in JavaScript, and offers them on every paid Claude plan, though Pro has to turn them on in /config.