Google Antigravity release notes: Agent Skills get their own codelab

Google Antigravity did not get a new public app version today. The changelog still tops out at 2.1.4 from June 11. But Google published a fresh practical codelab on June 17 about Agent Skills in Antigravity, and it has release-note value because it changes how teams can package repeatable agent work.

Antigravity is Google’s agentic development platform: a workspace where AI agents can read project context, propose changes, and use tools under human control. Agent Skills are small directory packages, usually centered on SKILL.md, that give the agent specialized instructions only when the task calls for them.

Source: Authoring Google Antigravity Skills, Google Codelabs, last updated 2026-06-17. Background: Google Antigravity Changelog.

Google Antigravity release notes: Agent Skills become practical

The new signal is not an app version like 2.1.5. It is an official Google guide that makes Skills much more concrete: what a skill is, how SKILL.md works, where skills live, and how Antigravity or Antigravity CLI can list them.

Google describes a skill as a directory-based package with a definition file (SKILL.md) and optional support resources such as scripts/, references/, templates, and static assets. The point is progressive loading: the model first sees metadata and loads the heavier working instructions only when the user’s task matches the skill.

For Nordic teams, this is more useful than another prompt collection. A good skill can carry your standard for release notes, code review, license headers, database rules, or customer-specific integration routines. You stop pasting the same long instructions into every new conversation.

Source: Authoring Google Antigravity Skills, sections "What is a Skill in Antigravity?", "The SKILL.md Definition File", and "Using Antigravity or Antigravity CLI".

What Google verifies in the codelab

Google shows two scopes for skills:

• Global scope: ~/.gemini/config/skills/, available across Antigravity, Antigravity IDE, and Antigravity CLI.
• Project or workspace scope: <project-root>/.agents/skills/, limited to one project.

The codelab also shows four examples: git-commit-formatter, license-header-adder, database-schema-validator, and json-to-pydantic. It says the user can ask Antigravity "What skills are available?" and that Antigravity CLI can use /skills to list installed skills.

This is where teams should slow down a little. I would not let an agent install global skills without review. Start with one skill in the project’s .agents/skills/, review SKILL.md, and let it run scripts only with clear approval gates, scoped permissions, environment variables or a secret manager for keys, and a log of what ran.

Source: Authoring Google Antigravity Skills, "Global Scope", "Project/Workspace Scope", "Installing the Skills", and the /skills example.

Short example: use the new Antigravity feature

Human step first: make sure Antigravity is installed, put a test skill in <project-root>/.agents/skills/, and list skills in Antigravity or Antigravity CLI using the method Google shows in the codelab.

Then paste this short agent prompt:

Inspect the skills available in this workspace and propose one project-scoped Skill for our release-note routine. Include a SKILL.md outline with name, description, trigger conditions, steps, constraints, and one verification check. Do not create or edit files yet; ask for approval first.

Good output should:

• separate installed skills from proposals that do not exist yet
• suggest project scope, not global installation, if the routine belongs to one repo
• write a concrete description that helps the agent trigger the skill for the right task
• end with a clear approval question before files are created or changed

What Hammer readers should test now

Do not test Skills by trying to automate everything at once. Choose a routine where you already have a human standard: release notes, proposal review, data validation, support triage, or integration checks.

Start with a skill that does not need secret keys. When it later connects to real systems, build the integration as a controlled Tool Forge routine: scoped API keys, a secret manager or environment variables, least privilege, output redaction, approval gates, and run logs. Then Antigravity becomes a workbench you can operate, not just another AI chat.

The next thing to watch in the Google Antigravity changelog is whether Skills get more visible version notes, UI management, or stricter policy controls. On June 18, the public app changelog still shows 2.1.4 as the latest entry, so this post is based on the new official codelab, not a new app release.

Source: Google Antigravity Changelog, checked 2026-06-18.