Google is testing ways to let its Gemini AI take direct action on items flagged in Gmail. Code discoveries point to an expansion of the existing AI Inbox that goes beyond surfacing to-dos. The assistant could soon start handling them. But it won’t act alone.

Strings unearthed in the latest Gmail app build reveal a new workflow. Users may see a “Start with Gemini” button next to suggested tasks. Tap it and the status shifts to “In progress.” Gemini works in the background. When it finishes, the item moves to “Ready for Review.” There the user checks the output, makes changes if needed, and decides whether to approve.

Sometimes the AI hits a wall. It then marks the task “Needs your input” and offers a “Give input” button. The human fills in missing details. This human-in-the-loop approach appears deliberate. One wrong automated reply to a client or boss carries real consequences. Google seems determined to avoid them.

The findings come from an Android Authority teardown of Gmail version 2026.09.28. Reporter Abhishek Yadav first flagged the work on X. His post showed screenshots of the new buttons and status labels. The code suggests Gemini can draft replies for some items instead of the full start button. That option reads simply “Draft Reply.”

A separate teardown by NerdsChalk uncovered related strings in an earlier build. One reads “Accomplish this Task.” Another handles “References: {URLS}.” The combination hints that Gemini may pull supporting links when it completes work. The feature remains hidden. No users have access yet. Rollout timing stays unclear.

This push builds on AI Inbox, which already scans incoming mail and highlights action items. It groups them into to-dos and topics. Some users love the clarity. Others find the suggestions noisy. The new agentic layer aims to turn passive recommendations into completed work. Yet it stops short of full autonomy. Approval remains mandatory before anything sends or marks complete.

And the idea fits a larger pattern. Google has spent the past year wiring Gemini deeper into Workspace. Earlier teardowns showed Deep Research gaining the ability to query Gmail and Drive directly. Custom agents for family coordination or finance tracking appeared in the Google app. A shared agent called Robin surfaced in group chats and email threads. All point to the same goal. Make the AI a participant that does more than answer questions.

Recent reports add context. On the same day as the Gmail findings, NerdsChalk detailed new source management tools in Google Chat. Users will add and remove reference materials. They can even attach a Drive presentation to set visual style. The controls suggest Google wants tighter oversight on what data its agents use. Limits on usage also appear in related code. Banners warn when approaching quota. Some features may sit behind paid tiers.

Executives have described the vision in public. At I/O 2026, Google introduced Gemini Spark, a 24/7 agentic assistant with native Gmail ties. It pulls facts from emails, docs, and sheets to draft updates or respond to customers. Small businesses, the company said, could let it watch inboxes so questions never go unanswered. Yet those demonstrations stayed high level. The APK strings show the gritty engineering required to make it safe for everyday use.

Privacy questions follow naturally. Gemini already reads inbox content for summaries and overviews. The new tasks require deeper access. It must interpret intent, locate related messages, perhaps edit drafts or update labels. Google says it does not train models on user data without permission. Still, the connector scopes for enterprise versions list read and modify permissions. Users must weigh convenience against the scope of access.

Early reactions on X mixed excitement with caution. Many professionals spend hours each week clearing inboxes. Automating the routine parts could free time for higher judgment calls. Others worry about errors compounding across long threads. A misread tone in one reply might trigger a chain of misunderstandings. The review step tries to catch those cases. Whether it proves enough will depend on how accurately Gemini interprets context.

Similar experiments run across the industry. OpenAI, Anthropic, and Microsoft all chase agentic systems that complete multi-step work. Google holds an advantage here. Its products already sit inside millions of inboxes and calendars. The data connections exist. The challenge lies in making the AI reliable enough that people trust it to act without constant supervision.

For now the code offers only hints. “View progress” and “Open task” strings suggest a dashboard where users monitor ongoing work. An “Active” label appears in related Robin agent code. These pieces imply a maturing framework. One that treats the AI as a teammate rather than a search tool.

Google has not commented on the findings. It rarely does before features reach beta. Yet the pattern of leaks across Gmail, Chat, Docs, and the core Gemini app shows clear momentum. The company wants its AI to move from helper to actor. The Gmail task system represents one careful step in that direction.

Success will hinge on execution. If the review flow feels frictionless, adoption could spread fast among power users and small teams. If it feels cumbersome, the feature may linger as an occasional novelty. Either way, the teardown signals that Google no longer sees the inbox as a passive archive. It views the inbox as a place where work gets done. And it intends to put Gemini in the middle of it.