Eric Zakariasson

His personal project list includes Scandinavian Design, an agent skill for restrained interface design, and a browser demonstration that lets readers switch between desktop and mobile examples. It also lists Web UI Bench, a comparison of models on the same interface applications, and cookbooks for Cursor and Grok. These projects make his interest in agent behavior concrete: how an agent is instructed, what environment it can use and how a person assesses the result.[1][5]

Supervising through a measurable goal

In “Human in the /loop,” Zakariasson describes giving a long-running agent a fixed definition of done before it starts. The target might be a test suite, a browser check or a performance measurement. The agent iterates against that target and stops when it succeeds, stalls or needs a decision. Notifications bring the person back at those points. His workflow moves some of the effort from continuously watching output to designing a check the agent can actually perform.[4]

Separate coordination from bounded work

“Let the smart model delegate” describes using one model to plan and coordinate while faster workers handle well-scoped subtasks. Each brief specifies a concern, context, a testable result and what to report. The coordinating model reviews the returned work. He also explains when to keep a problem in one session: a difficult judgment or tightly coupled investigation may not divide cleanly. The article is an account of his experiments with Cursor’s models and cloud agents.[3]