AI Agents Now Use Computers, But Infrastructure is Key

The question that once seemed like science fiction is now a reality: can AI agents actually use a computer? For years, the promise of AI assistants automating everyday digital tasks was hampered by their inability to reliably navigate interfaces, fill forms, or click buttons. Now, according to new data, computer-using agents are not only functional but are beginning to hold up in production environments at scale.

This advancement marks a significant step beyond simple demos. The models have improved dramatically, enabling agents to tackle narrow, repeatable workflows like updating systems of record, processing data through portals, and managing software where no clean API exists. The ability to deploy these capabilities to manage end-to-end tasks, previously requiring human intervention, is opening up new frontiers in business process outsourcing (BPO) and automation. As detailed in a recent analysis, the progress is tangible, moving the conversation from 'can they use a computer?' to 'can they reliably do this job?'

Beyond the Benchmark: Real-World Performance

The OSWorld-Verified benchmark, which tests an agent's ability to operate a real desktop across Windows, macOS, and Ubuntu, shows a dramatic leap. A year ago, the best-performing models scored around 42% on tasks. Today, the leading models, like Anthropic's Claude Fable 5, are achieving 85%, surpassing the ~72% score of human testers on the same tasks. This leap in capability is what has made production deployments viable.

However, benchmarks only tell part of the story. In real-world deployments, especially for standardized back-office work, the raw UI navigation is becoming a commodity. The durable advantage is shifting to the layers above the model: managing context, permissions, process knowledge, validation, and sophisticated error handling. Companies are finding that agents are brittle when work deviates from established protocols. For instance, if a retailer portal changes its layout overnight, an agent needs robust recovery mechanisms. The true test is not how many tasks an agent completes in a lab, but whether a business process can be reliably automated end-to-end.

The Infrastructure is the New Frontier

For enterprises adopting these technologies, the specific AI model executing tasks is often secondary. What matters is the surrounding infrastructure that ensures reliability, scalability, and security. Buyers are evaluating vendors based on their ability to demonstrate ROI, pass security reviews, and handle failures gracefully. One operator running millions of automated tasks monthly couldn't even recall which model was powering their system, indicating that the vendor's ability to swap models underneath without disruption is the true value.

This focus on infrastructure highlights a critical shift. The hard part is no longer getting an agent to click the right button. It's imparting the organizational context: the tribal knowledge, internal terminology, escalation paths, and precise validation rules that make work flow smoothly within a specific company. This specific, unglamorous knowledge is where focused startups can build a moat, rather than competing solely on model performance.

The Economic Case for Agentic Work

The economic argument for computer-using agents is becoming compelling. Running an agent on a frontier model can cost between $6-$8 per hour, with a range of $3-$15 depending on the harness design and how much work can be offloaded to deterministic code. This positions AI agents as roughly break-even against offshore BPO workers (around $10/hour fully loaded) and offers substantial cost savings, approximately 70-80% gross margin, compared to US-based back-office labor ($30-$45/hour fully loaded).

This economic advantage, combined with 24/7 availability and scalability, makes the case for adopting AI agents in areas previously dominated by human labor. The technology has moved from the research lab to production, proving its worth in automating the long tail of software interactions that lack APIs. The real inflection point is not just technical, but economic, paving the way for a new wave of automation in the real economy.