AI 程式工具
Antigravity CLI 1.1.6 Switches to Markdown-Defined Custom Agents, Bringing Tool Inheritance and Command Policies Under Version Control
Google Antigravity CLI adds a mechanism for creating main agents and subagents with `agent.md`, with frontmatter controls for MCP inheritance, visibility, and command execution policies. The new release also improves code search streaming and agent monitoring, though whether file-based policies can provide a reliable security boundary still requires testing.
Antigravity CLI 1.1.6 moves custom agents to a repository-friendly Markdown specification. Developers can create project-specific agents at `.agents/agents/<name>/agent.md` in a workspace or place them in the global configuration directory for use across projects. Each file consists of YAML frontmatter followed by a system prompt. Newly added fields include `mainAgent`, `subagent`, `hidden`, `inheritMcp`, and `commandExecutionPolicy`, which distinguish agent roles, determine whether an agent appears in the menu, control whether it inherits MCP tools, and restrict how commands may be executed. Dynamically created subagents are also written in the same format, reducing inconsistencies between version parsing in internal builds and external distributions.
This change allows agent configurations to undergo code review, follow branch-management workflows, and be shared across teams, making them better suited to large repositories than interactive settings stored only on individual users’ machines. The `/agents` panel can switch agents, display the status of background subagents, and terminate tasks. Switching agents within an existing conversation automatically forks the session to preserve the original context. Plugins can also bundle these agent definitions, enabling teams to package code-review, testing, or migration workflows as reusable components.
Version 1.1.6 also makes `/codesearch` display results incrementally as they arrive and adds support for cancellation with Esc, preventing searches in large repositories from blocking the interface. Agents now receive read access to the system temporary directory by default, reducing repeated authorization requests when processing build artifacts. However, this also expands the data surface visible by default, so enterprises should verify whether temporary directories might contain credentials or sensitive content left behind by other processes. Engineering teams should also confirm that `commandExecutionPolicy` is enforced by the execution layer rather than functioning only as prompt configuration, and test whether a malicious repository could elevate tool permissions by committing an `agent.md` file.