AI agents can pursue wrong directions for a long time before producing output. By the time you see the result, significant time and tokens are wasted. The alternative - exposing the agent's reasoning in real time - adds cognitive load and can feel noisy. The design question: when should users see the agent's thinking, and when should they be able to interrupt?

A background agent completes 90% of a task. The remaining 10% needs human judgment - a nuance the agent missed, a decision that requires context the agent doesn't have. If the handoff from "agent working" to "human editing" is clunky, you've just negated the benefit of automation. The transition is a design problem.

The "assistant" model -sitting in a sidebar, watching an agent work, ping-ponging back and forth -optimizes for the wrong thing. As models become capable of autonomous work for 45 minutes or more, the human becomes the bottleneck. The future of agent tooling isn't a better sidebar. It's a factory: spawn multiple agents, let them run, check in periodically.