At Google I/O 2025, the tech giant announced the latest advancements in Project Mariner, a next-generation AI agent that can perform automation tasks and navigate the web like a digital executive assistant. Integrated into the Gemini app’s new Agent Mode, the rapid innovation of AI-based automation, with AI agents and Agentic AI approaches marks a rapid evolution from traditional automation approaches.
“We think of agents as systems that combine the intelligence of advanced AI models with access to tools, so they can take actions on your behalf and under your control,” notes Google CEO Sundar Pichai on the Google I/O announcement blog.
Project Mariner: Google’s Web-Native AI Agent
Born out of research at Google’s DeepMind, Project Mariner aims to act directly on behalf of the user by understanding and manipulating web content, such as clicking, scrolling, and submitting forms just as a person would. It leverages advanced vision-language models to interpret not just text, but visual elements like buttons, charts, and images. This native multimodality, Mariner’s ability to process and act upon mixed media inputs, is a key differentiator in the growing landscape of AI assistants.
With the newly unveiled “Teach and Repeat” capability, users can demonstrate an action once, such as filling out a complex travel form, and Mariner will replicate the sequence across similar contexts. It can even handle multiple tasks in parallel, up to 10 simultaneous tasks in this latest release, making it capable of autonomously comparing and booking travel options or filtering data across web apps like Zillow and LinkedIn.
Agent Mode in Gemini: Intelligent Automation in Your Pocket
Google’s Agent Mode, now live in the Gemini app for Ultra tier subscribers, is the interface through which Mariner becomes truly useful. Tell Gemini to “find me two-bedroom apartments under $2,500 with in-unit laundry near downtown Austin,” and it will execute a search, apply filters, and surface listings, no clicks required. This task delegation interface blurs the line between search, automation, and conversation. The AI not only retrieves information via links, but it performs digital work on your behalf.
“This is a new and emerging area, and we’re excited to explore how best to bring the benefits of agents to users and the ecosystem more broadly,” shared Google on its blog.
Currently, Agent Mode is available in an experimental phase to subscribers of Google’s AI Ultra plan, with broader access anticipated in the coming months . Project Mariner is also expected to be accessible through Google’s AI Mode in Search, initially via the Search Labs program
Agentic AI vs. AI Agents: A Critical Distinction Emerges
As AI automation tools evolve, so too does the language used to describe them. Increasingly, experts distinguish between Agentic AI and AI Agents. While the terms were once used interchangeably, they now signal divergent design philosophies.
AI Agents are typically goal-directed programs designed to complete specific tasks, often relying on pre-coded workflows while Agentic AI implies a more cognitive, self-reflective system, one that adapts, plans, and reasons dynamically without hardcoded steps.
Mariner fits squarely in the Agentic AI camp. It doesn’t require you to map out a process. You simply describe your intent, and it figures out how to execute, even in unfamiliar digital environments.
This approach marks a sharp departure from platforms like Make.com and n8n, which empower users to build sophisticated AI and non-AI automation flows, but only if they understand logic trees, API endpoints, and error handling. These tools, while powerful, require a degree of technical fluency and upfront effort. Users must manually design each workflow: if X happens in App A, do Y in App B.
With Mariner, the AI handles that logic on the fly. Want to track invoice emails, extract totals, and update a spreadsheet? No need to draw a flowchart. Just ask. The difference is akin to building IKEA furniture versus hiring someone who walks in, reads your mind, and assembles the sofa before lunch.
AI-Powered Automation Heats Up
Likewise, traditional Robotic Process Automation (RPA) tools were designed to mimic repetitive human actions in structured environments, such as clicking buttons, reading fields, and moving data between legacy systems. But they break easily, require constant maintenance, and lack the situational awareness needed for more dynamic tasks. Now, RPA vendors UiPath and Automation Anywhere are working with Google via the Mariner API to evolve their offerings in this latest AI-powered automation age.
In contrast, Agentic AI systems like Mariner are robust against interface changes, can interpret new web layouts, and even troubleshoot errors mid-process. This agility makes them not only more powerful, but fundamentally more sustainable in fast-changing digital ecosystems.
Research from Randstad Enterprise notes that nearly 40% of RPA implementations stall due to the fragility of rule-based scripts. As AI agents become more autonomous and multimodal, organizations may no longer see value in maintaining brittle automation scripts when they can simply ask an AI to get the job done.
Other players are quickly moving into this space. OpenAI’s “Operator” initiative explores similar territory, while startups like Adept and Rewind.ai are racing to build “action-taking” AI agents. What sets Google apart is its broad ecosystem, from Search and videos to Chrome to Gmail to Android, which gives Mariner a head start.
