Google details Antigravity SDK for custom agent hubs

Google has detailed how developers can use its Antigravity SDK to build custom multi-agent control planes. The software uses the same runtime engine as Antigravity 2.0 and the Antigravity CLI.

The SDK is presented as an option for organisations that want to build their own agent hubs rather than rely on managed commercial platforms. It is aimed at developers running bespoke workflows or custom execution engines who need predictable runtime behaviour, full logging and sandboxed operation.

A multi-agent control plane is designed to monitor and manage large language model workloads. In Google's outline, that includes visibility into an agent's reasoning steps, tool calls and stored session state, along with an observability layer that can stream telemetry to a dashboard over WebSockets.

The example describes an operations engineer overseeing several active agents, including ones handling background research, document summarisation and task scheduling. According to Google, this type of work often leaves teams relying on fragmented console logs, manual inspection of JSON transcripts, and limited visibility into loaded skills, connectors, token usage and latency.

Core runtime

At the centre of the approach is the Antigravity SDK agent core, which manages model interactions, runs tools, generates reasoning traces and executes skills. A second component, middleware built around lifecycle hooks, intercepts events such as step starts, thought updates and tool calls, then relays them to monitoring systems.

Google says the runtime coordinates execution through five mechanisms: session initialisation and state attachment; skill resolution from filesystem paths; concurrent stream generation from a single model response; sandboxed execution of built-in tools; and telemetry interception through lifecycle hooks.

The session model stores trajectories, tool receipts and other artefacts under a root directory using a conversation identifier. This allows developers to resume prior context without an external database while preserving a record of multi-turn interactions for audit purposes.

Skills and tools

The SDK also uses a filesystem-based method for loading skills. Rather than relying on an in-memory registry, agents resolve domain-specific instruction bundles from directories containing a SKILL.md file, then add them to the system prompt.

Google describes this as a way to manage reusable instruction sets for tasks such as research or code review. In the example configuration, a local agent is assigned specialist skills by passing directory paths directly into the runtime.

The software also includes built-in file and workspace tools, including commands for listing directories, finding files, searching directories, viewing files, creating files and editing files. Declarative safety policies limit those actions to approved workspace directories and block filesystem operations outside authorised paths.

That workspace scoping is intended to address a central concern in enterprise agent deployments: limiting what an agent can access and change. Google says the policy model enforces those limits at runtime rather than leaving developers to write custom wrappers around file operations.

Telemetry layer

Lifecycle hooks form the monitoring and control layer around the runtime. Developers can use asynchronous hook functions to mark session start and end states, inspect tool parameters before execution, approve or block actions, and record tool results after execution.

In Google's example, these hooks send status and tool data to a live dashboard. That would let operators see whether an agent is running or idle, which tool it is trying to use, what arguments it passed, and whether the tool returned an error.

The same hook framework can also support human approval flows before a tool is allowed to run. That gives organisations a way to place an audit gate around sensitive actions without redesigning the agent itself.

Custom build route

Google's broader message is that enterprise agent adoption is splitting between buyers that want a managed platform and teams that prefer to assemble their own control layer. Antigravity SDK is positioned for the latter group, especially developers who want direct control over execution, observability and persistence.

Google links the SDK to model support, including Gemini 3.1 Pro and Gemini 3.8 Flash. It also says SDK-based agents inherit runtime updates automatically when the core engine changes.

For companies building internal agent hubs, the practical appeal lies less in the model itself than in the surrounding controls. The architecture Google outlines focuses on making agent sessions observable, resumable and restricted to defined workspaces, while giving operators one place to monitor several agents at once.

The dashboard model is intended to replace multiple terminal windows and manual transcript checks with a single operational view of agent activity, including "which Skills or MCP connectors are loaded for a given agent session" and "cumulative token usage and execution latency".