Mechanize Aims to Fully Automate Computer Work with AI Agents

San Francisco-based startup Mechanize is pushing the boundaries of AI office automation with an ambitious goal: to fully replace human computer work with AI agents. The company, founded by Tamay Besiroglu, Ege Erdil, and Matthew Barnett (formerly of research group Epoch AI), is building simulated digital workplaces designed to train AI agents through reinforcement learning.

Simulated Offices for AI Training

Mechanize's approach involves creating virtual workspaces that mimic real digital offices, complete with email inboxes, Slack, code editors, and browsers. AI agents perform tasks, receive rewards for success, and penalties for failure. "It’s effectively like creating a very boring video game," Besiroglu told the New York Times.

The founders believe this method will eventually produce agents capable of handling any computer-based job, though timelines vary. Barnett estimates 10–20 years, while Besiroglu and Erdil predict 20–30 years.

Beyond Coding: A Vision for Full Automation

Mechanize's ambitions extend beyond software development. The team envisions AI agents managing every digital task, from planning and communication to execution. "We’ll only truly know we’ve succeeded once we’ve created A.I. systems capable of taking on nearly every responsibility a human could carry out at a computer," the founders write.

However, the company offers few specifics on the social impact of widespread automation. While they support ideas like universal basic income, there’s no concrete transition plan. Barnett argues the mission is ethically justified if society becomes wealthier overall.

The "Bitter Lesson" and Reinforcement Learning

In an accompanying essay, Mechanize highlights the limitations of current AI systems, linking them to the "bitter lesson" of AI research: data- and compute-driven approaches outperform hand-designed algorithms at scale. The breakthrough, they argue, will come from massive-scale training in simulated environments.

This aligns with recent thinking from researchers like Sutton and David Silver, who advocate for agents that learn by doing, not just consuming human-written data.

From Coding Assistants to Generalist AI

Mechanize’s strategy combines human demonstration data with reinforcement learning in simulated offices. The goal is to create "drop-in remote workers" that delegate, plan, and fix mistakes like human colleagues. Current RL environments lack internet access and multi-agent collaboration, but Mechanize aims to build richer, more realistic training spaces.

Software Development: The First and Last Frontier

Software development is Mechanize’s initial focus, as it can be broken into discrete tasks. Yet, it’s also complex enough to serve as the ultimate test for agentic AI. While AI already handles code completion and testing, architecture decisions and team coordination remain challenges. Mechanize believes richer training environments could eventually automate even these roles.

Competitors: Major AI labs are also developing RL environments, from raw logs to simulated workspaces.

Key Takeaway: Mechanize’s vision hinges on scaling reinforcement learning in hyper-realistic digital offices—a gamble that could redefine the future of work.