Project Mariner (Google AI Agent) - First 5 Tests and Impression

Google’s Project Mariner browser agent delivers a mixed but promising first impression: it can reliably navigate the web, complete straightforward search-and-find tasks, and even execute simple code via an online Python runner—while running into hard limits around actions like sending emails and some interactive chat flows.

In the first test, Mariner successfully searched YouTube for the creator’s “Google Flow all about AI” video and returned a concrete metric: 25,000 views. The workflow looked like a typical agent loop—prepare a session, browse to the relevant site, run the query, then confirm completion—ending with an explicit “task complete” status and the view count displayed.

The second test targeted Gmail automation. Mariner could gather information about the latest Claude Anthropic live stream events from the web, but it stalled when asked to send an email. Even after the user manually logged into Gmail via takeover, the agent refused to complete the send step, producing a “cannot do that” style blocker. The result was a partial win: it could collect the live-stream details, but it couldn’t perform the final outbound action.

The third test focused on DeepMind’s diffusion model page and joining a weight list. Mariner found the correct page, clicked through to the “join the weight list” area, and handled cookie acceptance. It then reached a sign-in-gated form and successfully updated a field—changing a “profession” value to “engineer.” That demonstrated the agent’s ability to interact with multi-step web forms, though the overall experience was described as “far from perfect,” suggesting friction or brittleness in real-world flows.

The fourth test was the most impressive: Mariner was asked to find a way to test a small Python snippet online. It identified a site (v3schools) as a place to run code, initially crashed, then recovered after retry instructions. On the next attempt, it ran the code and produced the expected output: “The sum of seven and five is 12.” The agent required additional guidance (like excluding comments), but it still managed to locate an execution environment and complete the task end-to-end.

The final test attempted to converse with ChatGPT about the future of software engineers. Mariner struggled with the interaction layer: it navigated toward ChatGPT-related pages, but responses didn’t load, and an “internal error has occurred” message appeared when trying to run the prompt. The user ultimately abandoned the chat portion.

Overall, Mariner’s early strengths center on web navigation, search, and form completion, plus the ability to execute simple code through external tools. Its early weaknesses show up when the task requires privileged actions (like sending emails) or reliable interactive chat behavior. The takeaway is clear: the agent can be useful for research and browsing tasks, but it still needs guardrails and better reliability for action-heavy and conversational workflows.