Google Cloud Tech: Agent Harnesses Explained: Inside the Stack Behind Antigravity, Claude Code & Cursor

Try Google Antigravity → https://g.dev/cloud/antigravity-homepage
Explore Google Skills on GitHub → https://goo.gle/3UN4MWQ

What exactly is an agent harness, and why is it the secret to fully autonomous coding? If your coding agents are hitting a wall, the problem might not be the model you are using, it might be the stack. In this episode of The Agent Factory, we break down how to stop relying on manual prompting and start building environments where AI can dynamically solve complex software engineering challenges.

First, we sit down with Ryan La Popolo, the Google engineer who coined the term "agent harness". We will discuss why he hasn't manually written code since May, how to apply senior engineering principles to AI agents, and how agents are being deployed to operate Google Cloud. Then, Billy jumps into the code to demonstrate the difference between linear and closed-loop harnesses using the Agent Development Kit (ADK). Finally, Smitha reveals the three-layer stack you need to build high-performing agents today, featuring Gemini 3.8 Flash, Antigravity Boost, and the viral Google Skills GitHub repository.

Whether you are a developer trying to build your own agent loop, or a technical leader looking to scale AI autonomy across your infrastructure, you'll learn exactly how to configure the memory, tools, and guardrails necessary to make your agents reliable. You'll walk away with a clear understanding of when to use lightweight models for high-frequency loops, when to deploy orchestrators like Boost, and how to give your agents the precise domain knowledge they need to succeed.

Chapters:
0:00 - The Future of Agent Autonomy
0:30 - What Exactly is an Agent Harness?
2:22 - Ryan La Popolo on Coining the "Agent Harness"
5:00 - Shifting Left: Automating Context for Agents
10:30 - Do Agents Cohere Over Long Time Horizons?
16:47 - Building Agents for Google Cloud
19:44 - Code Deep Dive: Building Your Own Harness
22:45 - Linear vs. Closed Loop Agent Patterns
25:27 - 3 Layers of a High-Performing Agent Stack
26:40 - Why Gemini 3.8 Flash is Perfect for Agents
27:16 - Using Google Antigravity Boost
28:31 - Supplying Knowledge with Google Skills

More resources:
The Anatomy of Harness Engineering: How to Evaluate, Iterate, and Guard AI Coding Agents → https://goo.gle/4j3lCKU
Read Ryan's essay on Leveraging Codex → https://goo.gle/4inFlVv
Learn more about Gemini 3.8 Flash → https://goo.gle/3TuVfmT

Speakers:
Ryan Lopopolo:
GitHub → https://goo.gle/46T7sF0
hyperbola → https://goo.gle/4xIXqS5

Tilda Thurium:
X → https://goo.gle/4AbPQ4x
LinkedIn → https://goo.gle/4itBt5w

Smitha Kolan:
YouTube → https://goo.gle/Smitha-on-YouTube
LinkedIn → https://goo.gle/Smitha-on-LinkedIn

Luke Schlangen:
Website → https://goo.gle/4h1VC1r
LinkedIn →https://goo.gle/4iWWghS

Billy Jacobson:
LinkedIn → https://goo.gle/4cJk5WA
Medium → https://goo.gle/4ydgpVX

Watch more of The Agent Factory → https://goo.gle/4xi7qRL
🔔 Subscribe to Google Cloud Tech → https://goo.gle/GoogleCloudTech


Products Mentioned: Gemini, Gemini Flash, Gemini 3.8 Flash, Google Antigravity, Antigravity Boost, Google Cloud, Agent Development Kit (ADK), Google Skills, Firebase, Flutter, Google Maps, GitHub