Agentic AI Platform Change Log

Workflow orchestration

Google launched Antigravity Remote Control, a browser-first feature that allows developers to monitor, manage, and interact with long-running agent tasks across multiple machines. The update includes proactive push notifications to alert users when an agent requires human input to proceed.

This untethers developers from their workstations, enabling them to kick off complex, time-consuming refactors or test suites and manage them remotely. It significantly improves the usability of autonomous agents by solving the blind spot of walking away from a running task.

Security / enterprise

Google introduced new enterprise controls for Antigravity, allowing administrators to set strict boundaries on agent behavior. IT teams can now restrict file access to specific working folders, mandate human approval for terminal commands, enforce sandboxing, and limit web browser access.

These governance features are critical for enterprise adoption, ensuring that autonomous agents cannot inadvertently access sensitive data or execute destructive commands. Full audit logging and data privacy enforcement provide the necessary compliance guardrails for large organizations.

Agent capability

Google Antigravity added support for defining custom agents using standard Markdown format via an agent.md file. The release also introduces progressive streaming for the codesearch command, an optional index argument for copy, and default read access to the system temporary directory.

These enhancements give developers a more accessible way to configure agents and improve the speed and efficiency of searching codebases and managing files within agentic workflows.

Agent capability

Google added new controls for model reasoning effort by introducing a dedicated effort command and launch flag. The platform also introduced stable model slugs for consistent sessions, subagent model configuration options, and a redesigned model picker.

This allows engineering teams to dynamically balance latency and reasoning depth depending on task complexity while ensuring reliable model routing across multi-agent setups.

Security / enterprise

Security researchers disclosed a sandbox escape vulnerability where the Antigravity agent could write a file that a trusted external tool would execute to bypass the sandbox entirely. Google acknowledged the risk of these indirect prompt injections and rolled out patches to address the issue.

Security and compliance teams must verify their Antigravity environments are patched to the latest version to prevent malicious workspace files from compromising host systems during autonomous tasks.