Google Antigravity adds Dynamic Workflows for AI Pro and Ultra users
Google Antigravity now lets its coding agent write and run a Python script that directs a whole team of subagents. The feature, called Dynamic Workflows, arrived in version 2.32.0 of the Antigravity App on November 13, 2026, and is started with the /run-workflow command in both the desktop app and the Antigravity CLI. According to Google's documentation, it is available to subscribers of Google AI Pro and Google AI Ultra, which in the United States cost $19.99 a month and from $99.99 a month respectively.
The mechanism is simple to describe. Instead of handing a large task to one agent and hoping it keeps track of every step, the main agent writes a short Python program that defines the steps, decides which ones run in parallel and which run in sequence, and assigns each piece of work to a separate agent. Each of those agents runs in its own conversation and returns a result that has to match a JSON Schema, so file reads, test logs and intermediate reasoning stay out of the main conversation. The script can be saved to a .py file in the project and run again later.
Google released the feature one day after launching Gemini 4 Hydrogen, its faster model for coding and agentic execution, and four days after adding /swarm, a separate command for coordinating subagents over a shared message board. With Dynamic Workflows, Antigravity now offers six distinct ways to split work across agents, and only two of them, /run-workflow and /swarm, are restricted to the Pro and Ultra plans.
··········
HOW ANTIGRAVITY DYNAMIC WORKFLOWS RUN A TASK.
A workflow starts when the user types /run-workflow followed by a task, or describes the workflow in plain language. The agent reads a built-in skill named run-workflow, which tells it how such scripts are written, and then produces the Python program. Before anything runs, the user sees the script and approves or rejects it, unless the active permission preset already allows it. Rejecting the request and explaining what to change is the documented way to revise the plan.
Once approved, the workflow runs in the background while the user keeps working in the same session. When it finishes, the main agent receives the logs and the output of every phase and writes a summary. The individual workflow agents do not see the script and do not see one another, so the documentation instructs that each agent's prompt must contain everything that agent needs to know. That design keeps each subagent's context small and predictable, at the cost of forcing the main agent to be explicit about every handoff.
The scripts run in a separate Python process. Antigravity ships its own Python runtime, so users do not need to install anything, although the first run takes a moment while the bundled interpreter is unpacked. Developers who want a different interpreter can point the ANTIGRAVITY_PYTHON environment variable to a Python 3.8 or later installation. If security software blocks the bundled runtime, Antigravity falls back to a Python 3.8+ installation on the system path, and if none is found the workflow stops with an error explaining how to install one. Google also warns that on macOS the app does not read shell environment variables when it is opened from the Dock.
··········
THE BUILDING BLOCKS OF A RUN-WORKFLOW SCRIPT.
Google documents a small set of functions that the generated scripts use. They define what a workflow can and cannot express, and users who save scripts will end up reading and editing them.
........
Building block | What it does |
phase(name) | Starts a named phase that can be followed live |
log(...) | Adds a line to the current phase and to the main agent's result |
agent(...) | Defines an agent with a prompt, schema, role, type, workspace and model |
parallel(items, fn) | Runs items at the same time and returns results in order |
pipeline(items, fn) | Runs items one at a time, in order |
trajectory(...) | Returns earlier steps of the conversation, with optional filters |
args | Optional arguments passed to the script |
........
An agent can be one-shot, called once with await agent(...), or persistent, created once and then called repeatedly across turns. Persistence is what makes reviewer-and-fixer loops possible: a single reviewer agent can list problems, a set of fixer agents can work on them in parallel, and the same reviewer can then confirm that every problem it found has been addressed.
Three options control how each agent behaves. The model option accepts inherit, which is the default, or one of three tiers named flash_lite, flash and pro; an agent fails and lists the valid choices if the requested tier is not available on the user's plan. The type option selects the agent itself, a built-in read-only research agent, or one of the user's custom subagents. The workspace option keeps the agent in the current folder, places it in an isolated Git worktree with branch, or gives it a new working copy that shares repository storage with share.
One behavior deserves attention before running large jobs. If any item inside a parallel() call fails, the other items in that call are canceled. Google's documentation does not list numeric limits such as a maximum number of agents per workflow or a maximum runtime.
··········
WHO CAN USE ANTIGRAVITY DYNAMIC WORKFLOWS AND ON WHICH PLAN.
Google's slash-command reference lists /run-workflow as available on Google AI Pro and Google AI Ultra, in both the Antigravity App and the Antigravity CLI. The Dynamic Workflows page does not mention the free tier, and the plan restriction matches the one Google applies to /swarm. The table below combines the US prices from Google's subscription page with the Antigravity allowances described in the Antigravity plans documentation.
........
Plan | US price | Antigravity quota | Dynamic Workflows |
No AI subscription | Free | Weekly quota and weekly rate limit | Not listed |
Google AI Plus | $4.99/month | No Antigravity benefit listed | Not listed |
Google AI Pro | $19.99/month | Quota refreshed every five hours, weekly limit | Yes |
Google AI Ultra | From $99.99/month | Highest quota, refreshed every five hours | Yes |
........
Ultra is sold in two tiers in the United States, at $99.99 and $199.99 a month, with the higher tier carrying the highest agent rate limits Google offers. Pro and Ultra subscribers can also buy AI credits to keep working once their baseline quota runs out; the credits are consumed at Gemini Enterprise consumption pricing, and a setting called AI Credit Overages decides whether they are used automatically or never. Our guide to Google Antigravity pricing, free access and usage limits covers how those allowances evolved.
Google has not said how much quota a workflow consumes. Its plans documentation states that complex tasks consume more quota than simple ones and that limits may change to manage capacity, and a script that starts several agents in parallel performs the work of several agents. The model option is the practical lever here: routing routine steps to flash_lite or flash and reserving pro for the hardest steps is the documented way to control which tier each agent uses. Business and enterprise teams reach Antigravity under Google Cloud terms through Gemini Enterprise, and the Dynamic Workflows page does not describe availability for those accounts.
··········
DYNAMIC WORKFLOWS COMPARED WITH ANTIGRAVITY'S OTHER MULTI-AGENT TOOLS.
Antigravity now offers several overlapping ways to put more than one agent on a task, and Google's documentation includes guidance on choosing between them. The table below summarizes that guidance together with the plan requirements listed in the slash-command reference and in Google's Gemini 4 Hydrogen announcement for Antigravity.
........
Option | Best suited for | Plan requirement |
Orchestrator | One conversation with a frontier planner and a faster executor | Ultra (Carbon); Pro and Ultra (Argon) |
Subagents | Delegating one task while you keep working | Not specified |
Dynamic Workflows | Scripted control flow across many items or stages | Pro and Ultra |
/swarm | Peer agents coordinating on a shared message board | Pro and Ultra |
/boost | Deep reasoning on hard bugs with independent verification | Paid plans |
/teamwork | Agent teams for large, multi-milestone projects | Paid plans |
........
The distinction between Dynamic Workflows and the Orchestrator is the most useful one for everyday work. The Orchestrator, introduced on November 12 alongside Gemini 4 Hydrogen, decides on its own when to hand a task between Gemini 4 Carbon, which plans and explores, and Gemini 4 Hydrogen, which edits files and runs tests. Google reported that on Terminal-Bench 4.1 it reached close to Carbon's task success while using about 24% less quota and finishing about 19% faster on average. Dynamic Workflows works the other way around: the control flow is written down explicitly in a script that the user can read, approve, save and rerun. A workflow is the better fit when the same operation has to be repeated across many files or packages, when steps must happen in a fixed order, or when a team wants the exact same procedure to run every time.
The /swarm command, added in version 2.31.0 on November 9, sits between the two. Named workers and a verifier split a task and coordinate on a shared message board, which gives the agents more freedom than a script while still keeping a verification step. The /teamwork mode is aimed at the largest projects, and it is the mode Google Research highlighted when it offered six months of top-tier Antigravity access to 10,000 researchers.
··········
RUNNING AND MONITORING A WORKFLOW IN THE ANTIGRAVITY APP AND CLI.
Approval works differently in the two clients. In the Antigravity App, a prompt asks "Allow running this workflow script?" and offers four answers: allow this time, allow always in this conversation, allow always, or no. In the CLI, the prompt reads "Run this command?" with the options to run or cancel. In both cases an approval covers only that exact script, and a command-prefix rule cannot cover a Python script, so every new or edited script needs a new decision unless the user has chosen to always allow workflows. Each workflow agent also follows the user's permission preset and asks for approval in its own conversation when an action requires it. Stopping the parent conversation stops every running workflow agent.
In the app, a running workflow appears as a single step that shows its title, the current phase and an agent count such as "3/4 agents". Expanding it reveals the script, the phases, the logs, a card for each agent and the final output, and clicking an agent card opens that agent's conversation. The workflow also appears under Background Tasks in the side panel. In the CLI, a status row above the prompt shows the workflow, its phase and the agent count; Ctrl+O expands the phase tree, /tasks or Ctrl+T opens the task details, Alt+J jumps to an agent waiting for approval and Ctrl+K approves a pending request.
Google's documentation gives four example tasks that illustrate the intended uses. One reviews every Python file changed since the main branch in parallel with the flash model and logs a one-line verdict per file. Another has a persistent reviewer list the bugs in a billing module, assigns one fixer agent per bug in parallel and then asks the same reviewer to confirm the fixes. A third starts one agent per package in an isolated branch workspace to replace a deprecated helper function and run the package's tests. The fourth processes bug reports one at a time, with one agent writing a failing test and a second agent fixing the code until that test passes. To repeat a saved script, the user asks the agent to store it as a .py file and later runs /run-workflow with the path to that file.
··········
OTHER CHANGES IN ANTIGRAVITY APP VERSION 2.32.0.
Dynamic Workflows headlines a release that Google's changelog describes as containing 11 improvements and 13 fixes. The other additions include Everforest Light and Everforest Dark themes, keyboard shortcuts for moving between prompts and for submitting a prompt, stronger protections for browser tabs controlled by agents, and the ability to clone cloud GitHub repositories locally for remote agents. The fixes address browser compatibility, background processes left running after the app quits, sidebar filtering and text rendering.
The release follows a dense week for the product. Version 2.30.0 on November 9 added a Knowledge Work layout and the ability to import MCP servers, skills, commands, subagents and instruction files from Claude Code, Codex and Cursor; version 2.31.0 on the same day added /swarm; and version 2.31.2 on November 11 added Gemini 4 Carbon to the model selector for users signed in with Google Cloud application default credentials. Readers weighing the two paid tiers for this kind of agent work can compare them in our Google AI Pro vs Google AI Ultra comparison.
··········
GOOGLE ANTIGRAVITY DYNAMIC WORKFLOWS: CONFIRMED FACTS AND OPEN QUESTIONS.
The documented facts are clear. Dynamic Workflows shipped in Antigravity App 2.32.0 on November 13, 2026, it is invoked with /run-workflow in the app and the CLI, it is listed for Google AI Pro and Google AI Ultra, and it works by having the agent generate a Python script built from phases, one-shot or persistent agents, and parallel or sequential steps. Each script needs explicit approval, each agent returns schema-checked results in its own conversation, and agents can be assigned a model tier and an isolated Git worktree.
Several points remain undocumented. Google has published no figures on how much quota a workflow consumes, no maximum number of agents or runtime per workflow, and no statement on availability for free users or for Gemini Enterprise accounts. It has also published no benchmark for Dynamic Workflows itself, unlike the Terminal-Bench 4.1 figures it gave for the Orchestrator. For Pro and Ultra subscribers, the feature turns repetitive multi-agent jobs into scripts that can be reviewed before they run and reused afterward, while the cost of running them in quota remains something users will have to measure on their own projects.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
