Google shipped Custom Agents into Antigravity 2.0 and the Antigravity command line interface this month, letting developers define an agent’s role, tool list, and instructions inside a single markdown file instead of repeating the same guardrails in every chat session. The Antigravity IDE gets the same feature shortly after, per Google. The problem being solved is not glamorous: a general-purpose assistant that has to relearn a project’s testing conventions and dependency rules every session burns context and time before doing any real work.
Google’s own announcement concedes the limits directly. Custom agents do not replace skills or dynamic subagents, the two mechanisms Antigravity already ships for scoping what an agent knows. A markdown file that sets a role, a permitted tool list, and a system prompt is a configuration format, not a new capability. What changes is who decides the tradeoff between how much an agent is told and how many tokens that knowledge costs before it writes a line of code.
Google detailed the mechanics in a post on the Antigravity blog, its documentation and product site for the agentic coding line. Each custom agent lives as a markdown file with YAML frontmatter, saved to a project’s .agents/agents/ folder or to a user’s global ~/.gemini/config/agents/ directory. Committing the project folder to a repository means every teammate inherits the same specialized agents on checkout, with no manual setup. The frontmatter sets the model and the tool list; the markdown body underneath compiles into the system prompt at runtime.
A few mechanics go past the basic template. A mainAgent flag lets someone select a custom agent directly from the interface, rather than waiting for a coordinator agent to delegate to it as a subagent, which Google says other tools in this category do not allow. A commandExecutionPolicy field lets routine test and build commands run without approval while deletions and other high-risk actions stay gated behind manual sign-off. Nested lifecycle hooks can trigger a setup script before an agent starts and a verification script before any terminal command runs.
The pattern is now familiar across agent tooling. Anthropic ships a portable Agent Plugins bundle for packaging skills and MCP dependencies, Cursor pre-builds agent environments before a session starts, and Google is now shipping agent roles as version-controlled markdown files. Every vendor has separately landed on the same idea, put the agent’s configuration in a folder, and none has said whether these formats will read each other’s files or stay locked to their own tool.
Google’s post offers no measurement of the token savings or productivity gains it claims, only the mechanism and a code sample. That is consistent with a company blog describing its own feature rather than an independent benchmark, and it means the actual context-window savings for a real codebase remain unverified outside Google’s own telemetry.
Availability is narrower than the framing suggests: the feature works today in Antigravity 2.0’s desktop app and its CLI, with IDE support still pending. Teams already maintaining custom subagent definitions for Claude Code or Cursor should treat this as one more format to track rather than a reason to switch, at least until it is clear whether these role files travel between tools or stay fenced inside each vendor’s product.
Google’s Antigravity team detailed Custom Agents on the Antigravity blog, announced in August 2026.
