Source: Google Developers Blog — 2026-08-03
Summary
Google partnered with six independent open-source agent frameworks and tools — ADK, Agno, Browser Use, Eigent, Letta, and mem0 — to publish runnable, cloneable example agents that use Gemini 3 as the orchestrating model. Rather than another single-vendor demo, this is a cross-framework cookbook covering browser automation, multi-agent composition, enterprise workflow automation, and stateful long-term memory. The flagship example, a "Retail Location Strategy" agent built on Agno, composes specialized sub-agents that call Google Search and Google Maps plus a code-execution tool to analyze a market and generate a visual report autonomously.
Key Takeaways
- Six frameworks, six angles on the same problem: ADK and Agno both tackle multi-agent orchestration, but Browser Use and Eigent focus on web/enterprise automation while Letta and mem0 focus on giving agents durable memory across sessions.
- The Retail Location Strategy sample is the clearest artifact: a lead agent decomposes "should we open a store here" into sub-agents that pull live search results and map data, run code to analyze the numbers, and assemble a report — with no human stitching the steps together.
- The framing itself is notable: Google isn't just shipping its own agent stack, it's positioning Gemini 3 as the interchangeable orchestration layer underneath other people's frameworks — a bet that model choice, not framework choice, is becoming the differentiator.
- All examples are meant to be cloned and run, not just read — a rare move toward reproducible agent demos instead of static blog-post screenshots.
Reel Script
Hook (~18s, 38 words): Google didn't build one new agent demo — it shipped six, each built with a different open-source framework, all running the same model. That's a direct bet on which part of the agent stack actually matters.
Core Concept (~70s, 165 words): Here's the mechanism worth understanding. An "agent framework" like Agno or ADK is the scaffolding that decides how a task gets broken into steps, which tools get called, and how sub-agents talk to each other — think of it as the operating system for an AI agent, while the model itself is more like the CPU. Google's move here is to plug Gemini 3 into six completely different operating systems — ADK and Agno for orchestrating multiple sub-agents, Browser Use and Eigent for controlling a browser or enterprise software, Letta and mem0 for giving an agent memory that survives after the conversation ends — and show all six still work. The subtext: if the same model performs well no matter which framework sits on top of it, framework choice becomes a matter of taste, not capability, and the real competition shifts to whichever model is doing the reasoning underneath.
Hands-On (~65s, 155 words): The one worth sketching is the Retail Location Strategy agent, built on Agno. A lead agent gets a business question — "is this a good location for a new store?" — and instead of answering directly, it spins up specialized sub-agents: one queries Google Search for local market signals, another calls Google Maps for foot-traffic and competitor proximity data, and a third runs actual code to crunch the numbers into a chart. The lead agent then assembles those outputs into one visual report. Draw it as a hub-and-spoke diagram: one orchestrator at the center, three tool-calling sub-agents around it, converging into a single report node. That composition pattern — decompose, delegate, execute, synthesize — is the reusable piece here, not the retail example itself.
Takeaway (~22s, 50 words): If you're picking an agent framework right now, this tells you the framework matters less than you think — pick the one whose abstractions fit your team, because the model underneath is becoming swappable. Go clone one of the six examples and run it before you build your own from scratch.