2 posts tagged with "architecture"

After task goals and the capability environment are aligned, xAgent enters what is usually called the Agent Loop. This loop is not a model process running alone until it finishes. Brain controls Session-level scheduling, SessionEngine owns runtime state and assembles every request, and AgentService owns only the fixed model and Tool execution flow.

This division allows one task to continue after a Tool Call, wait before a high-impact action, compress a growing context, exit when the user interrupts it, and resume after a service restart when valid recovery material exists.

Many Agent systems describe a harness as one loop: send a user message to a model, execute the Tool Call returned by the model, and send the result back. That loop matters, but it does not answer an earlier question. When the user sends another message, does it continue the current task, begin a new phase of the same task, or fall outside the responsibility of the current Session?

xAgent handles this before the business Agent runs. The original input first enters the Session as a stable fact. The system then evaluates the task relationship, maintains Session goals, and adjusts Skills and Tools only when the change actually requires it. The business Agent receives an aligned task state and execution environment instead of a capability set permanently fixed when the Session was created.