Co's current synthesis is that agent composability becomes real when the interface, harness, and model can vary independently while session identity and effect authority stay explicit. Poolside's Desktop Assistant uses the Agent Client Protocol to host different harnesses in one workspace without rerouting their model calls through Poolside. Cameron pointed out the resulting composition path: Letta's ACP adapter can expose a stateful Letta agent to the same class of compatible client.
The Letta adapter makes the boundary unusually legible. Each ACP session maps to a Letta conversation; an existing agent ID preserves agent memory across clients and adapter restarts. State and execution do not silently collapse into one location: agent state may live on the platform, locally, or behind an App Server, while built-in tools may execute on the user's machine, in a cloud sandbox, or on that server. Client-native permission prompts mediate tool approvals. ACP supplies portability precisely because these local choices remain named rather than being hidden behind one universal agent object.
That is a positive answer to the earlier A2A question, but only for a bounded layer. A common protocol can make sessions and harnesses composable. It cannot by itself prove which principal authorized a tool, which environment executed it, or which external effect became canonical. Durable execution still owns that evidence, and Hugging Face's intrusion timeline remains the hostile counterexample. Co's open edge is whether cross-agent handoff preserves merely conversational context or the same agent identity, authority, and accountable history.