Antigravity's /run-workflow Has the Agent Write a Python Script to Direct Its Other Agents
Google's Antigravity 2.32.0, released November 13, 2026, adds Dynamic Workflows. The agent writes a short Python script that fans work out to parallel or sequential sub-agents, each returning a JSON-schema-checked result, while the user approves the script and follows each phase live. It is available on Google AI Pro and Ultra in the app and CLI.
Google’s Antigravity coding agent can now write the program that coordinates its own helpers. Version 2.32.0 of the Antigravity App, dated November 13, 2026 in Google’s changelog, adds Dynamic Workflows: instead of handling a large job inside one long conversation, the agent writes a short Python script that starts other agents, runs them in parallel or in sequence, and collects their results.
The feature is invoked with the /run-workflow command, or by asking for a workflow in plain language. Google’s documentation lists it for Google AI Pro and Google AI Ultra plans in both the Antigravity App and the Antigravity CLI. As with other Antigravity releases, the changelog notes that new versions roll out gradually and can take a few days to reach everyone.
From conversation to control flow
The design responds to a familiar problem with coding agents. When one agent reviews twenty files in a single conversation, every file read, search result and test log piles into the same context. Dynamic Workflows move that work into separate conversations. Each agent a workflow starts is a full agent with its own conversation, and it hands back a structured result that matches a JSON Schema. The main conversation sees the summary, not the clutter.
Google’s documentation describes the intended uses as tasks that split into many independent pieces, “one review per changed file, one audit per package,” or into staged steps such as writing a failing test and then fixing the code.
The scripts use a small set of building blocks. phase(name) groups work into named stages that appear in the interface. agent(prompt, schema, ...) defines an agent, and calling it once runs a one-shot job; keeping a reference to it lets the same agent continue across turns. parallel(items, fn) runs one agent per item at the same time and, per the docs, cancels the rest if one item fails. pipeline(items, fn) processes items one at a time. A trajectory() call lets the script read earlier steps of the conversation that launched it.
Each agent’s result passes through a finish tool that checks it against the schema, so the script receives parsed data rather than free text.
Choosing models and workspaces per agent
Workflow agents inherit the conversation’s model by default, but a script can ask for a lighter tier such as flash_lite or flash for high-volume checks, or pro for harder reasoning. The agent type can be the current configuration, the built-in read-only research agent, or one of the user’s custom subagents.
Workspaces are configurable too. inherit shares the current folder; branch gives each agent its own Git worktree, which the docs recommend when several agents edit code at once; share gives each a new working copy backed by the same repository storage. One example in the documentation starts an agent per package in an isolated branch to replace a deprecated helper and run that package’s tests.
Approval stays with the user
Antigravity shows the generated script and asks for approval before it runs, unless the user’s permission preset already allows it. An approval covers that exact script; the docs note that a command-prefix rule cannot cover a Python script. Agents launched by the workflow follow the same permission preset and ask for approval in their own conversations, where the user can open them from cards in the workflow step.
Scripts run in a separate Python process using an interpreter bundled with Antigravity, so no Python install is required. An ANTIGRAVITY_PYTHON environment variable points workflows at a different interpreter, version 3.8 or later. Users can also ask the agent to save a working script to the repository and rerun it later.
One of several ways to split work
Dynamic Workflows join a growing set of multi-agent options in Antigravity, and Google’s documentation includes a table on when to use which. Subagents delegate a single task. /swarm sets peers to coordinate on a shared message board. /teamwork runs a collaborative team on a large, multi-milestone project, starting with a scoping interview. The Orchestrator pairs a frontier planner with Gemini 4 Hydrogen for execution in one conversation.
Workflows sit at the predictable end of that range. Agents in a workflow “don’t see the script or each other,” according to the docs, so the script must give each one everything it needs. Google points users who need agents to negotiate interfaces or review each other’s diffs toward /swarm instead.
The same 2.32.0 release tightens agent-controlled browser tabs: cross-site frames now load signed out and stay hidden from the agent’s screenshots and page snapshots unless explicitly allowed, and blocked downloads are reported back to the agent.
- #Google Antigravity
- #Coding Agents
- #Multi-Agent
- #Orchestration
- #Python
Author
Raj M
Contributor
AI Systems Architect is a seasoned technology leader with over 15 years of experience in the IT industry working with Fortune 500 companies. With a solid foundation in multi-agent systems, open-source LLM infrastructure, and enterprise deployment, he excels at building scalable production-grade AI platforms.