After selecting Orchestrator (Carbon) or Orchestrator (Argon) in Antigravity CLI, Carbon or Argon handles read-only exploration, planning, and progress reviews, while Gemini 4 Hydrogen edits code, runs builds and tests, and follows the plan. If an issue falls outside the plan, control returns to the planning model. This is an automatic handoff within a single Agent session, not multi-Agent parallel execution with /swarm.
Suitable tasks: Routine development work that benefits from planning before implementation, such as feature development, bug fixes, refactoring, and debugging across multiple files.
Unsuitable tasks: Tasks that require multiple Agents to work in parallel or review one another. Use sub-Agents, Swarm, Boost, or Teamwork for those. For complex architecture work or cross-codebase migrations where every intermediate step must be handled by a high-reasoning model, select Carbon or Argon directly.
Supported model configurations: Orchestrator (Carbon) = Gemini 4 Carbon planning + Gemini 4 Hydrogen execution, and Orchestrator (Argon) = Gemini 4 Argon planning + Gemini 4 Hydrogen execution.
Supported client, Agent, or API: Google Antigravity CLI (interactive and headless modes).
Recommended reasoning tier and parameters: Orchestrator uses a fixed configuration. You cannot adjust it with --effort or /effort.
The following are the CLI selection methods given in the official documentation, not additional prompt templates.
# Start interactively and select the Carbon planner
agy --model "Orchestrator (Carbon)"
# Headless mode: use Argon for planning and Hydrogen for execution
agy -p "Add input validation to the signup form and run the tests." --model "Orchestrator (Argon)"You can also run /model in a CLI session, select Orchestrator, and press Enter. The documentation says this selection persists across sessions. Run agy models to see the model identifiers available to your account.
Confirm that your current Google account subscription includes the desired option: Carbon Orchestrator is available only with Google AI Ultra. Argon Orchestrator is available with Google AI Ultra, paid Google AI Pro, or eligible high-end-device trial plans. Neither Orchestrator is included with enterprise editions.
In the project directory, start agy --model "Orchestrator (Carbon)", or choose Argon based on the task. You can also select interactively with /model. For a single headless task, use agy -p "Task description" --model "Orchestrator (Argon)".
Provide a clear goal and acceptance criteria. The planning model first explores the codebase, resolves ambiguity, and creates a plan. Hydrogen then edits code, runs builds and tests, and verifies the results against the plan.
If Hydrogen encounters a blocker (for example, validation fails twice in a row, an interface behaves differently than expected, or a design choice was not covered by the plan), the Orchestrator hands control back to the planning model to revise the plan, then returns it to Hydrogen to continue. The planning model also reviews progress during long tasks.
Review the final workspace changes and validation results. Use /diff in the CLI to view the diff. Permission prompts are still handled according to the selected execution mode.
Google Antigravity Features (CLI tab): Explains that the CLI supports /model, and that within Orchestrator the planning model explores the codebase, creates a plan, and reviews progress while Hydrogen makes code changes and runs builds and tests. It also describes permission modes and default safety behavior.
Google Antigravity model documentation (Orchestrator): Lists the fixed division of work between Carbon/Argon and Hydrogen, subscription eligibility, dynamic handoffs, interactive CLI selection, and the agy --model and agy -p ... --model examples. It also states that Orchestrator does not support effort adjustments.
Specific permission boundaries: In the default request-review mode, the Agent requests approval before writing files, running Shell commands, or using network tools, except when an allow rule matches. The terminal sandbox is off by default. You can enable it with enableTerminalSandbox in settings.json or with --sandbox. The auto mode enables the sandbox and automatically approves workspace edits and commands inside the sandbox. The Agent asks for consent before accessing files outside the workspace unless allowNonWorkspaceAccess is enabled.
The official materials describe product capabilities and commands. They do not confirm that a specific account has access, and they provide no local execution results for this task. Check agy models and the model selector in the current account.
Model availability depends on the subscription plan. The availability of Carbon/Argon Orchestrator differs from the availability of Carbon/Argon base models selected on their own, so the latter does not establish access to the former.
/effort and --effort do not apply to Orchestrator. The documented fixed division of work for Carbon and Argon provides no parameters for customizing the planner or Hydrogen's reasoning tier.
This mode automatically hands work between planning and execution phases for the same Agent. For parallel delegation or cross-review, use the multi-Agent features listed in the documentation.
Keep permissions within the scope required for the task and review changes before accepting them. Do not use automatic execution authorization for unreviewed destructive actions, unauthorized access, or offensive purposes.
In an installed and authenticated Antigravity CLI, first run agy models to check available models. Then enter the target project and use one of the /model methods or startup arguments above. Submit a routine development task with clear acceptance criteria, observe the planning and execution handoff and any replanning, and use /diff to inspect the final changes. Actual availability and approval behavior depend on the account plan, project permissions, and CLI execution mode.
Gemini 4 Carbon