Testing AI agents for daily phone tasks over eight months

The limits of AI agents

Living with AI agents currently involves a mix of high success and total failure depending on the task. A user spent eight months handing over various phone chores to Gemini and Claude to see which ones actually saved time. The goal was to identify which errands could be fully automated and which required a human to stay in the loop.

The testing revealed that AI agents are currently best suited for rigid, predictable routines. For example, the user found that Gemini handles reordering a standard meal from DoorDash quite well. Because the order involves zero new decisions, the AI can navigate the app to fill a cart without human intervention.

However, the user noted that AI still struggles with tasks requiring an understanding of tone or nuance. They chose to triage their own inbox manually because they do not trust an AI to interpret the urgency of an email. This distinction between

Gemini and routine automation

Claude for file and calendar management

The cost of automation

Building smart home automations

Gemini has moved beyond simple one-at-a-time commands to a more agentic approach for home management. Instead of the user manually setting every trigger, they can now describe a general goal to the AI. Gemini then builds the automation by determining what should trigger the action, which conditions it needs to check, and the specific actions it should take to complete the task.

This shift allows the home to run on its own without constant human input. The user only needs to review the draft of the automation before saving it to the system. While this provides a hands-off experience, it is not entirely free, as building these types of automations requires a Google Home Premium subscription. Furthermore, any actions involving camera triggers require the more expensive Advanced plan.

Claude's file and calendar cleanup

Claude has proven to be a highly effective tool for organizing cluttered digital spaces like Google Drive. Rather than just suggesting folders, the AI can perform active cleanup tasks based on a single message. For instance, it can move old drafts into an archive or identify and send duplicate files directly to the trash, saving the user from manual sorting.

The AI also handles Google Calendar management with a level of ease that replaces traditional manual navigation. Users can ask Claude to move a meeting when plans change, and the AI updates the event for them automatically. This replaces a multi-step process of long-pressing, dragging, and hoping the event lands on the correct day, turning a complex task into a simple text request.

Finding your own AI balance

The effectiveness of these AI agents is not a universal rulebook that applies to every user. A person with a very messy Google Drive might choose never to trust an agent with their files, while someone with a different workflow might find different successes. The specific list of trusted tasks will vary based on individual needs and the level of trust a person has in the AI's accuracy.

The best way to determine what works is to experiment with specific chores and monitor the results. The goal is to see if the time saved by the agent is greater than the time spent double-checking its work. If the AI proves unreliable, the worst-case scenario is simply returning to the original method of tapping through the app manually.

Virtual windows and specific tasks

When the user asks Gemini to reorder a familiar meal, the AI operates by running the application in a virtual window. This allows the user to observe the agent navigating the interface and filling out the cart in real time. Because the AI is performing a known routine, the user can watch it complete the steps without needing to touch the screen themselves.

The user is careful not to use these automation requests for unfamiliar tasks, such as browsing a new menu. Since Google limits the number of requests available to users, spending one on an unknown order would be a waste of resources. Furthermore, giving the agent an open-ended job provides it with more opportunities to make incorrect guesses, which could lead to errors in the final order.

Searching for buried information

The user initially intended to use Gemini as a tool for finding specific details hidden within a large volume of personal data. This included searching for a specific figure buried deep inside a project document or trying to locate the date of a trip that was only saved as a photograph. The goal was to have the AI reason across different connected apps to pull out these scattered pieces of information quickly.

While Gemini was able to find some of these details, it often struggled with accuracy. The user noted that the AI would sometimes mix up information from two different documents while still sounding very confident in its answer. This forced the user to manually open the original files to verify the facts, sometimes requiring them to perform the search a second time to ensure the data was correct.