Google's latest announcements for agentic AI are more than just a new model. The release of Gemini 3.5 Flash and the Antigravity 2.0 development platform signals a shift from prompt-driven exploration to a more grounded, local-first engineering workflow for building agents. This matters because it changes the development loop from a slow, cloud-based iteration cycle to a faster, more tangible one on your own machine.
what changed: a fast model and a local orchestrator
Two main components define this shift. First is Gemini 3.5 Flash, a new model engineered for speed and efficiency in agentic workflows. It reportedly outperforms Gemini 3.1 Pro on most benchmarks while running significantly faster. This model is positioned as the high-speed engine needed for agents that must perform complex, long-horizon tasks with low latency.
The second, and more significant, piece is Antigravity 2.0. This is not just an API update; it's a standalone desktop application designed to be a central hub for agent interaction and orchestration. The platform is designed for developers to take an idea and build a production-ready application. This local-first approach allows for managing multiple agents in parallel, scheduling background tasks, and integrating directly with tools like Google AI Studio, Android, and Firebase.
why it matters: from prompting to engineering
For the last couple of years, building with large models has felt like working through a keyhole. You write a prompt, send it to a remote API, and get a response. Building agents required stringing these calls together with scripts and cloud functions. It worked, but it lacked the immediacy of traditional software development.
Antigravity 2.0 changes this dynamic. By providing a desktop application and a command-line interface (CLI), it treats agent development less like prompt engineering and more like systems engineering. The ability to orchestrate and deploy agents that can execute tasks in parallel from your local machine is a meaningful change. It encourages you to think about agentic systems as a collection of specialized workers, not a single monolithic model. This move from a simple request-response model to a managed, multi-agent system is where the real productivity gains will come from.
getting started with the new stack
Developers can access Gemini 3.5 Flash through Google AI Studio and what Google is calling Managed Agents in the Gemini API. The managed agent approach aims to remove the friction of infrastructure setup by delivering the power of the Antigravity agent harness through the API.
For local development, the Antigravity CLI provides a more direct interface. While the exact commands are still being documented, one could imagine a workflow for deploying a managed agent looking something like this:
# Fictional CLI command based on described capabilities
antigravity agents:deploy --name="daily-report-agent" \
--model="gemini-3.5-flash" \
--trigger="schedule --cron='0 9 * * *'" \
--task-file="./tasks/generate_report.json" \
--tools="google.workspace.sheets,google.workspace.docs"
This workflow, combining a powerful local CLI with cloud-managed execution, feels much closer to modern DevOps practices than the ad-hoc scripting that has characterized agent-building to date. Google AI Studio is also getting more integrated, with a new feature to export entire projects to Antigravity for local development and production deployment with a single click.
the takeaway for builders
This year's I/O updates are a clear signal that the infrastructure for building AI agents is maturing. The focus is shifting from the raw capability of a single model to the developer experience of building, testing, and deploying robust, multi-agent systems. We're moving from an era of AI that assists you to one where agents can independently navigate complex tasks across an entire workflow.
For engineers in the space, the message is clear: the tooling is catching up to the ambition. It's time to start thinking about agentic workflows not as a series of prompts, but as engineered systems that you can build and control from your own machine.
Top comments (0)