Orchestrator (Carbon) does not statically assign an entire task to one model in advance. Instead, Carbon and Hydrogen hand work back and forth at different stages within the same execution trace: Carbon explores and plans, while Hydrogen implements. If Hydrogen gets stuck, control returns to Carbon to inspect the state, revise the plan, and resume execution.
Suitable tasks: Long-running coding tasks that require first understanding repository dependencies and choosing an approach, then making routine code changes and running tests, and that may surface complex problems during implementation.
Unsuitable tasks: Tasks that require users to control model selection at every step, reproduce SDK call code, or rely on specific handoff conditions and tool configuration being publicly documented; the source does not provide those details.
Applicable model versions: Gemini 4 Carbon and Gemini 4 Hydrogen; the selector is named Orchestrator (Carbon) in the source.
Applicable client, agent, or API: Google Antigravity. The source also says the Antigravity SDK can explicitly combine multiple Agents and model roles, but does not show method names, request structures, or copyable code.
Recommended reasoning level and parameters: The source does not specify Orchestrator's internal reasoning level or parameters. When selecting Hydrogen on its own, the available levels are low, medium, high (the default), and xhigh; this setting is not a configurable Orchestrator parameter.
The following execution roles are distilled from the source. They are not a copyable system prompt or SDK configuration:
| Stage | Role | Responsibilities |
|---|---|---|
| Exploration and planning | Carbon | Explore dependencies in read-only mode, resolve ambiguities, produce a concrete plan, and specify when replanning should be triggered. |
| Implementation and testing | Hydrogen | Modify code, update packages, and run tests according to the plan. |
| Recovery when blocked | Carbon | Take over when Hydrogen encounters an obstacle, inspect the current state, revise the plan, and resume execution. |
flowchart TD
A[Carbon: Explore dependencies in read-only mode and clarify questions] --> B[Carbon: Create a plan and define replanning triggers]
B --> C[Hydrogen: Modify code, update packages, and run tests]
C --> D{Is execution blocked?}
D -- No --> E[Continue the current execution trace]
D -- Yes --> F[Hand back to Carbon: Inspect state and revise the plan]
F --> CIn the Google Antigravity model selector, choose Orchestrator (Carbon) and submit the full task to this orchestration mode.
Have Carbon inspect dependencies and task ambiguities in read-only mode first; make sure the plan specifies the implementation steps and when replanning is needed.
Have Hydrogen make code changes, update packages, and run tests.
If Hydrogen encounters an obstacle, hand control back to Carbon as described in the source. Carbon inspects the execution state, revises the plan, and resumes.
The source does not disclose how users configure or observe each handoff, nor does it define what counts as “blocked.” The steps above summarize only the roles and sequence explicitly described in the article.
The source describes Orchestrator as dynamic, bidirectional collaboration between Carbon and Hydrogen within a “single execution trace,” contrasting it with a static model router that estimates task difficulty from only the initial prompt.
The source's example has Carbon explore dependencies in read-only mode, clarify ambiguities, and create a plan with replanning triggers. Hydrogen then makes code changes, updates packages, and runs tests. If Hydrogen gets stuck, Carbon inspects the state, revises the plan, and continues.
Google reports that on Terminal-Bench 4.1, Orchestrator achieved a task success rate close to Carbon's, used 24% fewer average quota units, and completed tasks 19% faster. The source does not provide the sample size, intervals, or full test setup here. These are results reported by the source and should not be generalized as a guarantee for other tasks.
The source says the Antigravity SDK can explicitly control which models each role uses. It gives an open-source issue-triage and code-generation system as an example: the Hydrogen Agent handles issue triage, coding, codebase search, clarifying questions, and UI iteration based on PR comments; the Carbon Agent handles large-scale refactoring, new feature development, and security checks. The article provides no SDK sample code or call details.
Availability: The source says Orchestrator (Carbon) is rolling out to Google AI Ultra subscribers; at the time, paid Google AI Pro subscribers had access to Orchestrator (Argon). It also says enterprise users can use both in Gemini for business. Availability may change over time.
Orchestrator uses Carbon and Hydrogen as collaborative roles. It should not be described as a single Hydrogen model, or as a system where Carbon performs the entire task throughout.
The source does not disclose the internal routing algorithm, handoff thresholds, number of model calls, permission isolation, SDK APIs, cost calculation method, or materials needed to reproduce the experiment.
The reported 24% reduction in quota use and 19% speed improvement are only a summary of the source's Terminal-Bench 4.1 results. There is not enough detail to reproduce them independently, and they do not guarantee performance on a particular repository or task.
The SDK section supports only the claim that the SDK can explicitly combine Agents and model roles. To implement a working SDK pipeline, consult the official SDK documentation for the relevant version first; this article does not supply code that it does not show.
The original URL was opened and the article body checked through Tabbit on 2026-11-16. The page displayed the publication date, team byline, full article text, workflow description, performance claims, and availability. The page also links to Orchestrator documentation and the Antigravity SDK product page; this article does not infer SDK interfaces that the source does not show.
Gemini 4 Hydrogen