Project Mariner

» see deepmind.google/models/project-mariner as it lived, on the wayback machine

~ the obituary ~

this one sits close to what i do for a living. i work on vision models running on constrained hardware, where the whole discipline is deciding what the camera actually needs to see and how often. mariner was the same problem pointed at a browser instead of the ground, and it made the expensive choice on purpose.

google deepmind announced it on december 11, 2024. it was a chrome extension. you gave it a goal, it took a screenshot of the active tab, shipped that to gemini in the cloud, got back a decision about where to click or what to type, and did it. then it took another screenshot. it worked only in the tab you were watching, and it could not run in the background, which was framed as a safety property and was also an admission about how much you needed to supervise it.

the benchmark it launched with was webvoyager, which measures whether an agent can finish real tasks on real websites. mariner hit 83.5 percent as a single agent, best in class in december 2024, and around 90 percent with tree search. those numbers were genuinely good and they are part of why the ending is strange.

~ the case file ~

~ how it worked ~

the design choice underneath everything was pixels over apis. mariner did not integrate with sites, it looked at them. that is the correct move if you want an agent that can use the entire web on day one, because most of the web has no api and never will, and it is an enormous amount of compute to spend on reading a button that a dom query could have found for free.

at i/o in may 2025 google rebuilt it to run in cloud virtual machines instead of your local browser, which let it handle up to ten tasks at the same time while you did something else. it went to google ai ultra subscribers in the us, a tier priced high enough that the compute bill was not being hidden. the capability also started moving into the gemini api and vertex ai so developers could build on the same loop.

~ why it lost ~

screenshot-and-click is expensive per step and brittle in a specific way: a site redesign that a human absorbs without noticing can put the agent into a loop. every action costs a full round trip through a large model, and a twenty-step task means twenty chances to misread a modal dialog.

what overtook it was the coding agent. by 2026 the useful shape of an agent was one that writes and runs code against structured interfaces, calling the api where one exists and only falling back to looking at a screen when nothing else works. that is cheaper per task and far easier to verify, because the plan exists as text you can read. mariner was betting the web would stay unstructured. the agents that beat it just refused to accept the web on those terms.

~ what replaced it ~

on may 4, 2026 the mariner page stopped describing a product and started thanking people, with a line about the technology having voyaged to other google products. no blog post and no deprecation notice anywhere. for a research prototype that is defensible, though it had been a paid feature on the most expensive consumer tier google sells.

the capability did survive, which is the part the farewell line got right. gemini agent and the ai mode in search both carry pieces of it, and the browser-control work continues under other names. mariner was the demo that convinced a lot of people this category was real. it was also the version everyone stopped using once there was a cheaper way to get the same job done.

@}-->--

              )
             (_)
            .-'-.
            |   |
            |   |
            |   |
            |   |
          __|   |__   .-.
       .-'  |   |  `-:   :
      :     `---'     :-'
       `-._       _.-'
           '""""""

« previous  ·  random  ·  next »