Agents crossed 99% this year. Here is where the paying work goes next.

In June 2026, a browser-control agent achieved a remarkable 99.19% score on the WebVoyager benchmark, demonstrating its ability to handle 643 live-web tasks across 15 real websites. This milestone, announced publicly in October 2025, signaled the point at which agents can transition from mere demos to full-fledged staff-level capabilities. The author, having waited two years for this data, now predicts the direction of paying work in the AI industry.

Reliability, rather than model size, is the key differentiator. An agent that fails under 70% of the time is unreliable and unsuitable for client-facing tasks, while one achieving over 99% reliability can operate as staff, freeing humans from the need to monitor its output. The author emphasizes that the true value lies in the ability to automate tasks that previously required human intervention, such as dashboard creation, platform development, and interfaces built for human-only use.

The author's experience building AI agents for four years and selling over $1M in web and AI solutions highlights the importance of reliability and the pitfalls of relying solely on model upgrades. Instead, he advocates for focusing on the specification of the job, correcting agents in a standards document rather than in conversation, and narrowing the job scope to ensure success.

This approach allows agents to handle specific roles with defined outputs and schedules, leading to higher reliability and greater business value.

As the industry evolves, the focus is shifting from software writing to deployment and operation of AI agents. Enterprises are seeking Forward-Deployed AI Engineers who can work on-site with operations teams to build custom business agents. These engineers will be highly valuable, possessing a consulting-like skill set that bridges the gap between software development and business operations.

The Agent Ops Club, launched on July 15, 2026, offers a comprehensive operating model and a suite of pre-built systems for deploying and operating AI agents in real-world businesses. The license allows customization and deployment for client use, with the author taking no cut of implementation or managed operations services.