Antigravity Turns Agent Teams Into Reusable Python Scripts

Human Verified 0

Share to: , and

Abstract illustration of a developer platform orchestrating parallel coding agents through a glowing central control panel

Google has rolled out Dynamic Workflows in Antigravity, a new way to run multi-agent coding jobs as reviewable Python scripts instead of long, improvised conversations. Typing /run-workflow with a task makes the agent write a short script that coordinates parallel and sequential subagents, shows the script for approval, and then runs it in the background. The feature is live now in the Antigravity App and CLI for Google AI Pro and Ultra subscribers.

Abstract illustration of parallel coding agent tracks converging into a single verified workflow

How a workflow runs

When a developer types /run-workflow followed by a task, the main agent reads Antigravity’s built-in workflow skill and writes a Python script for the job. Antigravity displays the script and waits for approval before executing it, unless a permission preset already allows it.

The workflow runs in a separate process using the Python runtime bundled with Antigravity, so no local Python installation is needed. Developers who need a different interpreter can set the ANTIGRAVITY_PYTHON environment variable to point at it. The workflow keeps running in the background, so the developer can keep working in the same conversation while it progresses.

Each agent the workflow starts works in its own conversation and cannot see the orchestration script or its peers, so the script must supply every agent with its context, task and output format. Agents return structured results through a finish tool that validates each value against a JSON Schema.

File reads, searches and test output stay out of the main conversation, preserving its context window during high-volume jobs. The script organises work into named phases that can be followed live, with logs and agent cards grouped under each phase.

The parallel primitive runs one agent per item at the same time and returns results in order, cancelling the rest if one operation fails. The pipeline primitive processes items one after another instead, for jobs that touch shared files or test suites.

Scripts you can keep and re-run

A workflow script can be saved as a .py file in the workspace and run again later, so a recurring audit, review or refactor becomes a repeatable artefact rather than a prompt typed from scratch each time. Teams can review the script alongside their code.

Developers can pick the model per agent, using lighter tiers such as flash_lite or flash for high-volume checks and pro for deeper reasoning. They can also call on Antigravity’s read-only research agent or their own custom subagents for the job.

Parallel editing jobs can give each agent an isolated Git worktree through a branch workspace, so agents can modify files and run tests concurrently without overwriting one another. Dynamic Workflows carry no separate price and consume the plan’s existing model credits.

Google notes that high-fanout workflows can trigger many model calls, making model selection the main control over credit use. An approval covers only the exact script shown, so a changed script must be approved again before it runs each time.

Our Read

The telling shift is what becomes the unit of work. Until now, multi-agent coding in Antigravity was a conversation: the developer prompted, agents replied, and the plan lived in the chat history. Dynamic Workflows make the plan a script — something that can be read, approved, saved, versioned and re-run.

That turns agent teams from an improvisation into a repeatable process, which is what teams need before they trust agents with a whole branch. A script a developer can read before it runs is a strong answer to the control problem that has held agent coding back.

It also draws a clear line between Antigravity’s three coordination tools. Orchestrator, which arrived on 12 November, keeps one conversation and swaps models inside it, with a frontier planner handing routine execution to Gemini 4 Hydrogen. The /swarm command lets agents negotiate on a shared message board.

Dynamic Workflows are the deterministic option: scripted control flow, structured outputs, no peer-to-peer chatter. The trade-off is rigidity — agents cannot see each other, so the script must carry all context, and one failed item cancels the rest of a parallel batch.

Why It Matters

For development teams, the practical win is repetition. The same lint-style review across every changed file, the same package audit, the same test-then-fix loop can now be saved and re-run, with a reviewer agent checking the fixes afterwards on every run.

Google is betting that the coding-agent race will be won not by the smartest single model but by the most controllable team of them. Rivals building agent platforms now have to match the bundle — and the script, not just the model.

Background

Antigravity is Google’s agent-first development platform, and the pace of additions has been rapid. Google shut down Jules on 29 October as Antigravity gained a plugin marketplace, then added /swarm and Orchestrator within days of each other this month. Dynamic Workflows complete the set with the scripted, repeatable option.

Source: Google Antigravity docs and AlphaSignal

Share to: , and

Faheem R. Khattak

Senior automation engineer and AI marketing journo covering LLMs, algorithms, analytics, core updates, Bing search, and technical SEO.

0 Comments

Leave a Reply

Optional Take Tag:

🛡️ Spam protection: Honeypot + Time-Trap Active

This site uses Akismet to reduce spam. Learn how your comment data is processed.