agentLoop function

AgentEventStream agentLoop({ required List<Message> prompts, required Context context, required AgentLoopConfig config, required StreamFunction streamFunction, required ToolExecutor toolExecutor, CancelToken? cancelToken, })

Starts an agent loop with new prompt messages.

The prompts are appended to context and announced with MessageStartEvent/MessageEndEvent; the loop then streams the model, executes tool calls via toolExecutor, and re-invokes the model until it stops calling tools, fails, or cancelToken is cancelled. Provider failures never throw (errors-as-events): they end the loop with an assistant message whose AssistantMessage.stopReason is StopReason.error or StopReason.aborted.

The returned AgentEventStream replays all events to a late listener and its EventStream.result resolves with the messages produced by the run.

context is copied defensively; the loop mutates only its own copy.

Implementation

AgentEventStream agentLoop({
  required List<Message> prompts,
  required Context context,
  required AgentLoopConfig config,
  required StreamFunction streamFunction,
  required ToolExecutor toolExecutor,
  CancelToken? cancelToken,
}) {
  final stream = AgentEventStream();
  _drive(
    stream,
    runAgentLoop(
      prompts: prompts,
      context: context,
      config: config,
      emit: stream.push,
      streamFunction: streamFunction,
      toolExecutor: toolExecutor,
      cancelToken: cancelToken,
    ),
  );
  return stream;
}