怎样搭建真正可用的团队代理：五项设计决策，比多接一个模型更重要

Team agents can address knowledge bottlenecks and broken handoffs, but the design should start from the shared work problem rather than from a preferred software platform.

A team can maintain shared knowledge and skills while keeping individual agents private, preserving personal style without giving every user the same access or conversational history.

Expert, common-work, bridge, and chief-of-staff agents solve different coordination problems, so each archetype requires a different emphasis on knowledge, standards, permissions, or ongoing learning.

A shared agent is the wrong choice when individual judgment matters more than standardization, nobody owns its knowledge, or coordination costs exceed the work it removes.

A team agent needs role-specific responsibilities and an explicit list of prohibited actions, especially around commitments, interpersonal disputes, and movement of information between private and shared spaces.

Choose the simplest deployment that two colleagues will actually use, while explicitly deciding who can read their conversations and where the agent stores newly learned information.

Shared knowledge should be collected from its owners, reconciled for contradictions, explicitly approved, and maintained over time so that one person's outdated assumption does not become the team's default truth.

Access control must cover whose credentials the agent uses, who may request work, where the answer is delivered, and which human request caused each recorded action.

A durable team agent needs a named owner, clear engagement rules, accessible records of team decisions, and repeatable tests that reveal when its behavior has drifted.

Knowledge curation and agreed working rules can remain valuable when tools change, so teams can start with their existing stack instead of prematurely rebuilding a general-purpose agent platform.