Agent 正在成熟：模型更强，更强调升级机制

In business, I see guidelines for agents increasingly focused on keeping humans in the loop. It feels paradoxical: the talk used to be about agents taking work from humans; now it is about fitting them into human workflows. Ironically, agents entering corporate life face the same challenges as human employees. When generative AI became popular, we saw a surge of creative output: text, pictures and videos. But AI labs have always looked towards corporate automation as a source of revenue. Humans can be as creative as they want, but inside a bureaucratic corporation, success is not only about skill. It also requires following The Process; here, creativity means knowing where to stop and when to escalate. But putting a human in the loop is easier on an architecture diagram than in real life. Anthropic says that Claude Code users approve 93% of permission requests. This is why it built Claude Code auto mode: after enough interruptions, people stop evaluating requests carefully and simply click “approve”. A mechanism designed to provide oversight starts producing the opposite result. Approval is not escalation: the first asks a human to confirm the agent’s chosen action; the second hands them a situation the agent cannot safely resolve within its authority. A useful escalation should give the human enough context to understand why the agent stopped and what decision is required. Instead of offering only “approve” and “deny”, it should show the impact of the available options and allow the human to redirect the work or pass it to someone with the right authority. Google’s agentic design patterns and rationale for ADK 2.0 point in the same direction. Use deterministic workflows for known business logic and agents where ambiguity requires reasoning. Bring in humans for judgment, responsibility or policy exceptions. Before escalating, a system should run routine checks, retry recoverable failures and seek a safer route. Anthropic’s auto mode does this: when an action is blocked, Claude can try another approach; only repeated denials stop the process and bring in a human. The goal is not to remove humans or ask them to approve more actions, but to use their attention better. As agents improve, they should handle routine work independently and interrupt people less often—with better-prepared cases. The future is not endless permission dialogs, but agents recognising the limits of their authority and bringing humans only the cases that require judgment or responsibility.