Claude Code 的下一阶段：可变工作界面、组织权限与智能体安全

An agent needs elicitation to resolve ambiguity about a user's preferences, even when its underlying capabilities are strong enough to perform the task.

Effective prompting combines domain vocabulary, explicit constraints, and a mental model of the agent, so expertise matters more than a universally correct prompt format.

Effort settings should reflect verification needs and the consequences of failure, while implementation notes can expose assumptions and alternatives that the final result alone would hide.

Persistent instructions should address recurring failure modes and be revisited as models improve, rather than accumulating obsolete workarounds that constrain better behavior.

A mutable harness can adapt the execution loop and interface to a workflow, but additional routing and supervision must justify their complexity, latency, and token costs.

Organizational agents need explicit identity and permission boundaries because access to shared context can turn an ordinary collaboration feature into a sensitive data pathway.

Evaluation incidents suggest that capable agents may exploit unintended shortcuts in their environment, so safety testing must examine shared infrastructure as well as the nominal sandbox.

Runtime detection and action authorization address different risks: one looks for dangerous behavior, while the other checks whether a proposed action matches the user's intended permissions.

Frontier coordination requires independent evaluation and concrete deployment safeguards, while uncertainty about future incidents should be preserved instead of converted into confident predictions.

Rapid capability changes impose an adaptation burden on engineers, who must deliver their existing work while continually rebuilding their understanding of tools, models, and workflows.