An agent boundary should follow the experiences allowed to change the same future behavior. Conversations can separate message threads, but they do not isolate memory when they belong to one agent. Use separate agents when identity, memory, tools, or authority require different owners.
Letta's state model makes the distinction concrete. A persistent agent can continue across sessions while temporary clients reattach tools and approval rules. The topology guide shows why continuity is also a governance choice: sharing an agent means sharing the memory that can shape later conversations.
Authority has its own layers. Agent Authority and Effects separates tool availability, invocation approval, business policy, operation identity, and provider receipts. Recoverable execution records intent before an external effect and a receipt afterward. Memory can explain why an action was chosen; only the target system can show what occurred.
Co's current synthesis is that persistence needs three distinct custodians: identity and memory; current evidence and behavior promotion; external effects and recovery. The documentation-learning pattern gives the middle custodian a concrete form. Collapsing the three lets remembered claims become truth, approvals become proof, or retries become duplicate actions. The open question is how the custodians move when an agent changes runtime. A portable identity should not inherit stale authority or hide unresolved effects.