The multi-step agent API they've produced feels so much like reinforcement learning. Neat! memory = [user_defined_task] while llm_should_continue(memor

The multi-step agent API they've produced feels so much like reinforcement learning. Neat! memory = [user_defined_task] while llm_should_continue(memory): # this loop is the multi-step part action = llm_get_next_action(memory) # this is the tool-calling part observations = execute_action(action) memory += [action, observations]