Want your agent to write better code with fewer tokens? Ask the Google AI Team about Agent Skills!

Last week @stephr_wong, Kevin Hou, and Andy Zhang explored the Antigravity Editor and Agent Manager.

This week, @greggyb is back with special guests Luke Schlangen and Joe Hanley to dive into Agent Skills for Gemini CLI.

Ready for your agent to write better code with fewer tokens? The team will explore Agent Skills like the recently released Agent Skills for Firebase with the help of Gemini CLI. They'll test just how far these skills can take us.

Here's what's on deck:

• Exploring Agent Skills and how they improve code generation
• Testing the newly released Agent Skills for Firebase
• Pushing the limits of what Agent Skills can do with Gemini CLI

Stream Details

• 📅 When: March 3, 9 AM PT / 12 PM ET
• 📺 Where: We'll be simulcasting this livestream right on the DEV homepage!

Subscribe or Get Notified on YouTube

Drop your questions in the comments below. What do you want to know about Agent Skills? What are you curious about? They'll answer as many as they can live on the stream.

Comments

[Winter]

I've heard much about MCP servers previously and Agent Skills more recently. Could you provide clarity on:

• How do Agent Skills differ from other agent-based tools, such as MCP Server?
• What advantages do Agent Skills offer over MCP Server?

[Jess Lee The DEV Team]

@_winter_1314 just wanted to let you know that the Google AI Team selected your question! Right around this timestamp:

[Ryan]

How does the Agent Manager in Antigravity differ from the standard Gemini CLI when it comes to running multiple skills in parallel? Can they talk to each other or share context?

[Wei He]

What are some methods and tools to evaluate the quality and effectiveness of an agent skill?

[Wei He]

Will using subagents help with token usage? Do they have access to Agent Skills too?

[Wei He]

Do agent skills come with version control? Could you provide some best practices and tips on managing agent skills?

[Jean-Sebastien Beaulieu]

Yes — I would treat an agent skill like code, not like a loose note.

A practical minimum is: keep it in Git, add a semantic version in its metadata, maintain a short changelog, freeze the tool/schema contract, and test both positive and negative cases. The negative tests matter most: does the skill refuse unsupported actions, avoid the wrong tool, and stay within its declared scope?

I also version the executable tool or “motor” separately from the SKILL.md. Otherwise, an instruction change and a behavior change become impossible to diagnose.

I wrote a deeper breakdown here: dev.to/jsb-securedme/real-plugins-...

The short rule is: a skill teaches; a tool acts. Version and test both layers independently.