@Nylorun

The session's files: results too large to show, files tools return, files of earlier turns, and files the agent saves.

Asking whoever gave the task: a question, or an app to connect. The turn waits for the answer.

The Agent's memory: Markdown files under /memories that outlive the session, with their index pinned at the head of the conversation.

The agent's plan, kept with the task and shown to whoever gave it, as A2A progress.

The agent's computer: a shell and file tools in its sandbox, under the network policy of its runtime.

Triggers the agent sets up itself: follow-ups in this conversation and routines in a session of their own, kept by the Agent.