Antigravity 2.0

Antigravity introduces a dedicated environment for orchestrating autonomous agents, moving beyond ad-hoc script calls to a managed, UI-driven command center. The shift means that teams no longer need to stitch together separate scripts for each agent; instead they can launch and supervise them from a single dashboard. This change reduces operational overhead and makes it easier to track agent interactions across multiple initiatives.

The platform adds dynamic sub-agents that can be defined on the fly, allowing a single high-level request to be split into concurrent subtasks. By distributing work automatically, engineers can achieve faster turnaround on complex problems without manually managing thread pools or job queues. The benefit is a more scalable approach to problem solving, though developers must still design the decomposition logic that drives sub-agent creation.

Scheduled Tasks bring cron-style automation into the agent world, letting agents run periodic checks or maintenance jobs without manual triggers. Once a schedule is set, the agents execute in the background and report progress through Artifacts, which act as structured deliverables. This reduces the need for external schedulers but requires that teams define appropriate cron expressions and handle any failure modes within the agent logic.

Customization is exposed through downloadable skills, MCPs, and JSON hooks that can be scoped globally or per workspace, giving developers control over agent behavior. Projects group conversations and settings, supporting custom permissions that help isolate work across teams. Implementing these extensions will involve creating or importing skill definitions and configuring hooks, which adds upfront development effort but yields a more adaptable automation layer.

Live voice transcription leverages Gemini Audio models, turning spoken prompts into text that agents can act upon, and the software is distributed for both Apple Silicon and Intel architectures. Engineers must install the appropriate binary for their hardware and ensure access to the Gemini models for transcription to function. The solution is therefore bounded to supported Google-provided runtimes and may not operate on other hardware or without the associated audio model services.