wjgoarxiv/antigravity-swarm

Overview

This skill dispatches autonomous sub-agents to perform coding and project tasks inside the Antigravity IDE. It supports manual single-agent dispatch and a dynamic Auto-Hiring mode that builds and runs a team for complex missions. Plan Mode confirmation is built in to avoid accidental usage of API/model limits.

How this skill works

You can run a single sub-agent with dispatch_subagent to handle a focused job, or call run_mission which uses a Planner to generate a subagents.yaml team configuration and an Orchestrator to execute them. Sub-agents communicate via a file-based protocol and a simple CLI shim that recognizes write-file and run-command tokens, enabling automated file edits and shell execution. The orchestrator runs agents in parallel threads/processes and tracks progress in shared memory files (task_plan.md, findings.md, progress.md).

When to use it

• Offload isolated jobs like writing a test, analyzing a directory, or generating a specific file.
• Coordinate a multi-step app or feature that benefits from specialized roles (UI, backend, tests).
• Quickly prototype by auto-hiring a team from a high-level mission description.
• Run long-running background agents and monitor via JSON logs in an IDE agent workflow.
• Retry or isolate failing components without restarting an entire pipeline.

Best practices

• Write clear, self-contained task descriptions for single sub-agents to avoid ambiguity.
• Use Plan Mode for expensive or high-API-cost missions; confirm generated teams before execution.
• Prefer --format json and log files when running agents from an IDE agent so you can poll and visualize status.
• Limit shared-file contention by keeping small, well-scoped writes and using progress.md for status reports.
• Use gemini-3-pro or gemini-3-flash models as recommended to ensure file-shim operations work correctly.

Example use cases

• Create a small utility: dispatch a sub-agent to add a CLI script and tests.
• Build a feature: run_mission to auto-hire UI, API, and test sub-agents for a Todo app.
• Refactor a codebase: dispatch specialized agents for dependency analysis, code changes, and test updates in parallel.
• Continuous integration helper: spawn agents to run linters, unit tests, and generate coverage reports and aggregate results into progress.md.