Google Launches Custom Agents in Antigravity 2.0 for Specialized AI Coding Workflows
Summary
- • Google introduces Custom Agents in Antigravity 2.0 and CLI, enabling specialized, role-specific AI coding assistants per project
- • Custom Agents are file-based YAML/Markdown configs with scoped instructions, tools, model settings, and permission constraints
- • Version-controllable agent files committed to .agents/agents/ automatically become available to all teammates on checkout
- • Feature extends dynamic subagents — main agents can delegate to custom sub-agents with their own model choice and permissions
Details
Custom Agents in Antigravity 2.0
Google introduces Custom Agents as a first-class feature in Antigravity 2.0 desktop app and CLI; Antigravity IDE support follows shortly after the initial launch.
YAML/Markdown file-based format
Custom Agents use Markdown files with YAML frontmatter, stored at .agents/agents/ (project-scoped) or ~/.gemini/config/agents/ (user-global), following the same pattern as Skills.
Solves specialization and context bloat
General-purpose assistants require re-explaining project conventions every session; Custom Agents pre-encode role-specific rules and toolsets, keeping context lean and maintaining consistent, predictable behavior.
Version-controllable team AI workflows
Committing agent configs to a repository's .agents/agents/ directory makes them instantly available to all teammates who check out the code, standardizing AI-assisted workflows without manual setup.
Extends dynamic subagent delegation
Main agents can delegate work to Custom Agents specifying model selection and permission levels, building on the dynamic subagent framework launched earlier in the Antigravity roadmap.
Progressive discovery support
Like Skills, Custom Agents support progressive discovery — agents determine whether to load a full agent config based on the task at hand, minimizing token overhead across large agent configurations.
Product launch details for Google Antigravity 2.0 Custom Agents, sourced from the official Antigravity blog via TLDR AI newsletter.
What This Means
Google's Custom Agents feature in Antigravity 2.0 marks a meaningful shift in AI coding tools from single-model assistants toward multi-agent team architectures, where complex projects are divided among role-specialized agents rather than handled by a single general-purpose system. The version-controllable, file-based design means AI agent behavior becomes a formal part of software projects — reviewable, shareable, and standardized across teams — rather than living in individual chat sessions that disappear between conversations. For engineering organizations, this lowers the barrier to systematic AI adoption and enables consistent AI-assisted workflows that survive team member turnover. It reflects a broader industry trend of treating AI agents as configurable, persistent team members with defined roles and authority levels.
Sources
- Introducing Custom AgentsAntigravity