TestingCatalog reported finding an unreleased toggle inside Google’s Gemini desktop app that switches between an Ask mode and an Assign mode, a split the outlet says mirrors the Chat and Work split in OpenAI’s Codex. Nothing here has shipped. TestingCatalog says it located the feature through app teardown work, not a Google announcement, and the distinction matters because Assign is not another chat surface.
According to TestingCatalog, Assign behaves like Spark, the cloud agent Google introduced in May, but with wider capabilities layered on top. The build the outlet examined lets Assign scope itself to a folder: a default location, one used in a prior session, or a folder picked for a single task. That constraint is meaningful if it ships, since it caps what an autonomous mode can reach on a person’s machine rather than leaving it unbounded.
TestingCatalog also found an option to connect one Gemini desktop instance to another computer running the app. The outlet says this could enable remote control between machines, but it could not determine how the connection authenticates or what access it grants once established. It expects the mechanism to resemble Codex’s approach, a conclusion TestingCatalog draws from the code it reviewed rather than from any Google statement.
Two integrations stood out in the same build. One ties into a computer use skill TestingCatalog spotted in earlier Gemini desktop testing, letting a tagged task operate local apps directly. The other connects to Obsidian, the note taking app that stores files as local Markdown vaults. TestingCatalog says it cannot separate the two readings: a wider knowledge-base ambition for Gemini, or Google engineers simply using Obsidian themselves while they build something equivalent. A separate Customize tab, also unreleased, is meant to hold skills, connectors and plugins, though the version TestingCatalog saw offered little beyond apps already built into Gemini.
Code traces cited by TestingCatalog point toward compatibility work for a next major model, which the outlet says may be Gemini 4, and an updated image model it identifies as Nano Banana 2.5 Flash. Google has not confirmed either name. Code referencing a model internally does not guarantee a release date, a final name, or that the model ships at all.
That caveat applies to the whole set of findings: TestingCatalog notes every feature described here remains in closed testing with no public release timeline, and unreleased app code is routinely altered or cut before launch. A desktop assistant that can be assigned a task, scoped to a folder, and given a channel to another machine is a different security proposition than a conversational chat window. It needs permissioning, audit logs and a defined boundary on what “assign” can touch, and that governance question is the part worth tracking as these features surface rather than the feature list itself.
Teams evaluating Gemini as a desktop agent platform should treat Assign mode, the cross-machine option and the Obsidian connection as directional signals, not commitments, and hold production planning until Google confirms scope, permissions and a ship date.
Reported by TestingCatalog (Bedros Pamboukian and Alexey Shabanov), published September 5, 2026.
