The future of AIAgent won't just stop at adding an automation layer to humans, but will gradually evolve into a true system operation layer. Many capabilities that seem important today, such as skill, MCP, and CLI, are essentially interface forms of a transitional period.

Their existence is to let models learn to call the external world first, but in the long term, these capabilities will be absorbed internally like drivers in an operating system, becoming the Agent's instinctive actions. By then, people won't care about what protocol is used to connect, which commands were called, or how many intermediate layers were involved—people only care whether the objective is completed stably.
A truly mature Agent shouldn't expose tools as the primary interaction object. CLI has too high a cost for humans; it's more suitable for machines than for the general public. The future focus is not on teaching people more commands, but on letting the Agent itself understand, choose, orchestrate, and self-correct. Tools will become increasingly invisible, capabilities will be increasingly internalized, and interactions will be increasingly outcome-oriented.
So many people say GUI user interfaces are finished, but I don't agree with that. It doesn't mean interfaces will disappear. On the contrary, the stronger the Agent, the more it needs a high-density, low-burden GUI panel to help humans/machines quickly understand system status.
Even AI shouldn't read all feature documentation before operating; instead, it should first see the global status and system operation: which tasks are currently running, which resources are occupied, which objectives are completed, where risks exist, and why the system made these decisions. The best Agent experience in the future won't be command-line manipulation or chat-style exploration, but rather a collaborative sense that is viewable, interventional, and trustworthy. Whoever can handle both hidden complexity and clear global awareness simultaneously will be closer to the next generation of Agent.