Google is adding a new feature, Orchestrator, to Antigravity, its AI-powered development environment, to help enterprises improve software development efficiency by automatically coordinating multiple Gemini models, including the newly introduced Gemini 4 Hydrogen.
The new feature, which allows developers to delegate planning, coding, and review tasks to different models within a single coding session, combines a planning model, such as Gemini 4 Carbon or Argon, with Hydrogen, which the hyperscaler describes as a lighter execution model.
While the planner models explore the codebase, identify dependencies, and develop an implementation plan, Hydrogen handles execution, including editing files, writing code, and running builds and tests, the hyperscaler wrote in a blog post. That execution differs from conventional model routing, which typically selects a model based on a prompt and assigns it the entire task.
Orchestrator, instead, transfers control between the planning and execution models as work progresses, with the planning model first exploring the codebase, identifying dependencies, resolving ambiguities, and preparing an implementation plan before handing control to Hydrogen for execution, the hyperscaler explained in the feature’s documentation.
Hydrogen hands control back to the planner if it encounters problems, such as repeated test failures or unexpected behavior, allowing the latter to reassess the execution and revise the plan before Hydrogen resumes implementation, according to the documentation.
These handoffs, the hyperscaler said, happen automatically within a single response, without developers having to manually switch models or configure the transitions.
The new Orchestrator’s ability to use different models for different tasks in a single coding session is likely to boost productivity for developers, according to Pareekh Jain, principal analyst at Pareekh Consulting.
“This is like pairing a senior architect with a superfast coder. This could improve coding speed, debugging, code quality, and completion of complex tasks,” Jain said.
Primarily because developers often abandon tasks when AI coding agents struggle with complex requirements, repeated test failures, or unexpected code behavior, echoed Amit Kumar Jena, AI development head at IT consulting firm Kanerika.
The bigger gain, however, could be in token efficiency, which would help developers and CIOs control AI spending, Jena pointed out.
“Developers who rely on a single, more capable model for every task risk spending more than necessary on routine coding work, while using a lighter model throughout could compromise performance on complex tasks. By automatically assigning work to models based on the reasoning required, Orchestrator could help reduce unnecessary model usage and make AI coding costs more predictable,” Jena said.
For CIOs, particularly those overseeing large engineering teams, this could provide greater control over AI spending and help them scale AI-assisted software development without quickly exhausting usage limits or driving up costs, Jain echoed. However, coordinating multiple models within a single development workflow could complicate accountability, auditing, and security for enterprises, according to Jain.
“CIOs need visibility into which model made a particular decision or code change. Without model-level logs and clear attribution, enterprises could find it difficult to audit AI-generated code, investigate errors, establish accountability when something goes wrong, or understand why a particular model was selected for a task,” Jain said.
In addition, Jena pointed out that Antigravity’s documentation does not currently describe per-step logs identifying which model performed each action.
Enterprises, therefore, should test independently whether they can trace individual code changes to the model that generated them and the implementation plan it followed, he said.
Security, too, could be a concern.
Mainly because each model introduces its own behavior and potential risks when accessing source code, secrets and production credentials, Jena said.
“Enterprises, therefore, should give coding agents dedicated identities, restrict credentials to the specific tasks they perform, and require human approval before code changes are merged,” he added.
Beyond access controls, enterprises should also examine how each model handles their intellectual property, including where prompts and source code are processed, what data is retained, and whether their enterprise agreements cover every model the Orchestrator can invoke, the analyst further said.
Despite these concerns, developers who want to use the feature can access Orchestrator through the model selector below Antigravity’s conversation prompt box, the company wrote in the blog.
Currently, the hyperscaler offers Orchestrator with Carbon as the planner model to Google AI Ultra subscribers, while Orchestrator with Argon as the planner is available to both Google AI Pro and Ultra subscribers.
The selected configuration remains active throughout the conversation until developers choose another model, it wrote.
Developers, the hyperscaler added, also have the option of selecting Gemini 4 Hydrogen separately as a singular option for tasks that don’t require a planner model in Antigravity.
Gemini 4 Hydrogen, which has a 1 million-token context window, is priced at $1.50 per million input tokens and $7.50 per million output tokens, it further wrote.