News

Gemini Adds Mid-Task Steering for Active Agent Runs [APK Teardown]

Google App code suggests Gemini could let you steer an active agent task with new information, showing when it adjusts its approach, though the feature is not yet visible in the app.

Google App code suggests Gemini could let you steer an active agent task with new information, showing when it adjusts its approach, though the feature is not yet visible in the app.

0 Comments

What to know

  • Google App code adds feedback for steering Gemini while an agent task is running.
  • The status messages show Gemini pivoting, recalibrating, and updating its approach after a user update.
  • We found the unreleased code in Google App v17.70.10.sa.arm64, but could not surface the behavior in the app.

Gemini’s agent-processing interface already describes longer-running work through messages such as “Working on it,” “Using Chrome auto browse,” and “Running tasks in parallel.” That broader agentic auto-browse capability is already public, and the underlying processing interface has been present in the Google App’s code since June 23, 2026, or earlier.

Now, a teardown of Google App v17.70.10.sa.arm64 reveals an important update: Gemini is adding feedback that indicates you can give it new information while an agent task is underway, prompting it to change direction without beginning an entirely separate task.

About APK teardowns — An APK teardown works by reading the not-yet-shipped code inside a beta build, which lets us preview features a developer is working on before they’re announced. But work-in-progress code can change, stay hidden behind a server-side flag, or be scrapped before it ever reaches your phone.

Gemini can adjust an agent task while it is running

The new strings, seen below, point to a steering flow inside Gemini’s existing agent-processing interface. Each message acknowledges that an update has been received and describes an immediate change to the running task: “Pivoting with that in mind,” “Updating the approach,” and “Recalibrating with your update,” among others.

assistant_robin_agent_processing_steered_connecting_1 = “Pivoting with that in mind”
assistant_robin_agent_processing_steered_connecting_2 = “Adjusting based on that”
assistant_robin_agent_processing_steered_connecting_3 = “Got it, shifting gears”
assistant_robin_agent_processing_steered_connecting_4 = “Factoring that in now”
assistant_robin_agent_processing_steered_connecting_5 = “Updating the approach”
assistant_robin_agent_processing_steered_connecting_6 = “Course-correcting with that”
assistant_robin_agent_processing_steered_connecting_7 = “Taking that into account”
assistant_robin_agent_processing_steered_connecting_8 = “Recalibrating with your update”
assistant_robin_agent_processing_steered_connecting_9 = “Switching tracks with that”
assistant_robin_agent_processing_steered_connecting_10 = “Understood, pivoting now”

Together, the messages describe more than routine progress reporting. They indicate that Gemini can accept an update during active processing and reflect that change back to the user as it revises its approach. That could make longer agent runs easier to guide when a requirement changes or additional context becomes relevant.

This extends Gemini’s existing processing interface

This is an update to an existing feature, not the start of a new agent interface. Earlier Google App builds already included controls for expanding and collapsing Gemini’s worklog, along with status text for thinking, working, browsing with Chrome, and running parallel tasks.

assistant_robin_agent_processing_collapse = “Collapse”
assistant_robin_agent_processing_expand = “Expand”
assistant_robin_agent_processing_show_all = “Show all”
assistant_robin_agent_processing_show_less = “Show less”
assistant_robin_agent_processing_thoughts = “Thinking it through”
assistant_robin_agent_processing_working = “Working on it”
assistant_robin_agent_processing_worklog_title = “Using Chrome auto browse”
assistant_robin_agent_processing_parallel_tasks = “Running tasks in parallel”

The new pivot-status messages add a clearer feedback layer to that established interface, specifically for moments when a running task changes course in response to the user.

This steering update is not available to users yet, and we’ll watch future Google App builds for signs of its rollout.

Leave a Reply

Your email address will not be published. Required fields are marked *