TL;DR: An interactive study by the New York Times put autonomous AI agents to the test in a laptop environment with pre-configured apps. The results reveal that while agents excel at programmatic tasks, they suffer from severe UI navigation loops and formatting limitations, casting doubt on immediate corporate replacement strategies.
Behind the Experiment: Testing AI Agents on Office Jobs
The tech industry is witnessing a significant corporate push to replace human employees with autonomous AI agents. To test the validity of these corporate expectations, researchers deployed AI agents to act as office workers, giving them full access to a laptop with pre-configured applications. The experiment aimed to answer a fundamental question: Can artificial intelligence actually do an office job?
While autonomous agents, which act independently based on detailed instructions, excelled at tasks they could solve by writing computer programs, they struggled with human language nuances and browser-based user interfaces.
Detailed Task Breakdown: Slack and Google Sheets
The first test assigned to the AI agent was to gather feedback from colleagues via Slack and input that data into a Google Sheets spreadsheet. The "colleagues" were simulated using Slack bots that answered (or did not answer) using pre-written prompts. The agent was instructed to ask colleagues to rate a recently completed all-hands meeting on a scale of one to five and record the scores.
The agent's performance was a mix of brilliant automation and clunky workarounds:
- UI Avoidance: The agent initially avoided the browser interface entirely. It used Claude's built-in connection to search for the first colleague on Slack.
- API Exploitation: After switching to Slack in Chrome, the agent got an attendee list. Instead of messaging people individually, it programmatically gathered their user IDs to contact them via an API. It successfully sent the rating request to nine employees simultaneously.
- Error Resilience: The agent handled curveballs well. When an attendee's name was misspelled as "Sara Johnson" instead of "Sarah Johnson," the agent resolved the error and messaged the correct person. It also bypassed a non-responding attendee without getting stuck.
- Data Entry Limitations: To create a spreadsheet, the agent entered "sheets.new" into the browser address bar. However, it failed to format the data, writing entire rows of information into single cells. Unable to solve this in Chrome, it copied the data to the clipboard and pasted it in.
- The Rename Loop: Renaming the spreadsheet triggered a severe loop. The agent repeatedly appended its new title to the default text, resulting in "Meeting Feedback Meeting Feedback Meeting FeedbackUntitled spreadsheet." It eventually accessed the "File > Rename" menu and executed custom JavaScript to set the title before saving it to Google Drive.
In subsequent tests, the agent was asked to suggest staff cuts. While it completed the task, it made an important error. In a final task to fill out 17 forms, the agent blazed through the tedious parts but struggled with a trivial step.
The Human Cost: AI as a Catalyst for Layoffs
Despite the clear limitations demonstrated in these tests, corporate executives are moving forward with aggressive workforce reductions. Data from Layoffs.fyi shows that more than 200 tech companies have cut roughly 120,000 jobs in 2026. Meta, Oracle, and other major technology firms have executed substantial cuts, citing artificial intelligence as the primary driver. Following the layoff of approximately 1,100 employees, the chief executive of Cloudflare openly stated that he expects AI to replace workers in middle management, finance, and marketing, representing a massive shift in corporate strategy.
Key Takeaways
- Programmatic Excellence: AI agents are highly effective at writing programs and using APIs, successfully managing multi-recipient Slack communication and resolving minor errors.
- UI Navigation Failures: Agents struggle with basic user interfaces, falling into infinite renaming loops and failing to format simple spreadsheet cells without workarounds.
- Critical Execution Errors: In more complex tasks, such as suggesting staff cuts or filling out web forms, the agents made important errors or struggled with trivial steps.
- Premature Layoffs: Despite these technical bottlenecks, tech companies have laid off over 120,000 workers in 2026, targeting administrative and management roles.
