Cloud Agent architecture: scheduling and executing AgentRun
A cloud Agent needs to keep working after a request leaves the browser, report progress, and preserve the result. The platform separates user interaction from execution and gives each execution a durable AgentRun identity.
Different responsibilities, one AgentRun model
The service layer accepts a request, checks its user and work scope, and creates an AgentRun. The distributed scheduler dispatches and executes each AgentRun according to its authorized scope and execution policy. Conversation, scheduled, and background work share this model; they differ in how work starts and is managed.
Durable state coordinates instances
AgentRun state and results are stored outside an individual service process. That gives multiple service and scheduler instances a shared view of which work is queued, running, or complete. Ownership of an active AgentRun is checked when progress and final results are written, so a late response from a former scheduler instance cannot replace a newer terminal state.
Progress is separate from the final result
Progress events help the interface show what is happening while an AgentRun is active. They are delivered through an event channel; the durable AgentRun record remains the source of truth for the final state. Conversation history and deliverables are kept with their authorized work scope so people can return to the work later.
Define behavior when execution stops
A distributed design must say what happens if an executing process disappears. The platform uses time-bounded ownership of an active AgentRun; an AgentRun whose ownership expires can be marked failed for review. It is not silently assumed to have succeeded or automatically resumed from the interrupted step.