AI Agentçèªæè¿åï¼å 认ç¥ä¸åææºå¶çå®ç°å½AI Agentä¸åä» ä» æ¯æ§è¡é¢å®ä¹ä»»å¡çç¨åºï¼èæ¯è½å¤å®¡è§èªèº«è¡ä¸ºãåæé误并ä»ç»éªä¸æç»è¿åçæºè½ä½ï¼è¿æ å¿ç人工æºè½ä»å·¥å ·åèªä¸»æºè½è¿åºäºå ³é®ä¸æ¥ãæ¬æå°æ·±å ¥æ¢è®¨å 认ç¥ä¸åææºå¶å¨AI Agentä¸çå®ç°åçãæ¶æ设计ä¸å®æ代ç ãä¸ãå¼è¨ï¼ä¸ºä»ä¹Agentéè¦å 认ç¥è½åä¼ ç»AI Agentçå±éæ§å¨äºå ¶å³çå®å ¨ä¾èµäºè®ç»é¶æ®µå¦ä¹ å°ç模å¼ãä¸æ¦é¢å¯¹è®ç»åå¸ä¹å¤çå¤æåºæ¯ï¼Agentå¾å¾æææ çã人类ä¹æ以è½å¤ä¸æéåºæ°ç¯å¢ï¼æ ¸å¿è½åä¹ä¸å°±æ¯å 认ç¥ï¼Metacognitionï¼ââ对èªèº«è®¤ç¥è¿ç¨ç认ç¥ä¸çæ§ãç±»æ¯äººç±»çå¦ä¹ è¿ç¨ï¼å 认ç¥çæ§ï¼å¦ç解é¢æ¶æè¯å°èªå·±å¡ä½äºï¼éè¦æ¢ä¸ªæè·¯åæè¯ä¼°ï¼å¤çèè¯éé¢ï¼æ»ç»è§å¾çç¥è°æ´ï¼æ ¹æ®åæç»ææ¹åå¦ä¹ æ¹æ³å°è¿å¥æºå¶ç§»æ¤å°AI Agentä¸ï¼ä½¿å ¶è½å¤ï¼èªæè¯ä¼°ï¼å¤æå½åè¡å¨çè´¨éåæææ§é误è¯å«ï¼ä¸»å¨åç°èªèº«å³çä¸ç缺é·çç¥ä¼åï¼åºäºåæç»æè°æ´æªæ¥çè¡ä¸ºæ¨¡å¼æç»è¿åï¼éè¿è¿ä»£åæå½¢ææ£åæ¹è¿å¾ªç¯è¿ç§è½å对äºæ建çæ£èªä¸»ãå¯é çAI Agentè³å ³éè¦ãäºãå 认ç¥æºå¶çæ ¸å¿æ¶æ2.1 å 认ç¥ç³»ç»ä¸å±æ¨¡åä¸ä¸ªå ·å¤å 认ç¥è½åçAI Agenté常å å«ä¸ä¸ªæ ¸å¿å±æ¬¡ï¼| å±æ¬¡ | å称 | èè´£ | |------|------|------| | L1 | æ§è¡å±ï¼Executorï¼ | æç¥ç¯å¢ãæ§è¡å ·ä½å¨ä½ãå®æä»»å¡ | | L2 | çæ§å±ï¼Monitorï¼ | å®æ¶è¯ä¼°æ§è¡å±è¡ä¸ºï¼æ£æµå¼å¸¸åç¶é¢ | | L3 | åæå±ï¼Reflectorï¼ | 深度åæåå²è½¨è¿¹ï¼çææ¹è¿çç¥å¹¶æ´æ°æ§è¡å± |è¿ç§åå±æ¶æçæ ¸å¿ææ³æ¥æºäºè®¤ç¥å¿çå¦ä¸çå 认ç¥çæ§ç论ãæ§è¡å±è´è´£åï¼çæ§å±è´è´£è§å¯ï¼åæå±è´è´£æè为ä»ä¹ã2.2 å 认ç¥å¾ªç¯çå·¥ä½æµç¨class MetaCognitiveAgent: def __init__(self): self.executor Executor() # æ§è¡å± self.monitor Monitor() # çæ§å± self.reflector Reflector() # åæå± self.memory EpisodicMemory() # ç»éªè®°å¿åº def metacognitive_loop(self, task): å 认ç¥ä¸»å¾ªç¯ï¼æ§è¡ â çæ§ â åæ â è¿å episode [] # é¶æ®µ1ï¼æ§è¡ä»»å¡ while not task.is_completed(): # æ§è¡å±çæå¨ä½ action self.executor.decide_action(task.state) # æ§è¡å¹¶è§å¯ç»æ result task.execute(action) episode.append((task.state, action, result)) # é¶æ®µ2ï¼å®æ¶çæ§ signal self.monitor.evaluate(action, result, task) if signal.confidence 0.5: # 触åå³æ¶åæ self._interrupt_and_reflect(episode) # é¶æ®µ3ï¼äºå深度åæ reflection self.reflector.reflect_on_episode(episode) self.memory.store(episode, reflection) # é¶æ®µ4ï¼çç¥è¿å self.executor.update_policy(reflection.improvements) return task.resultè¿ä¸ªå¾ªç¯ç¡®ä¿äºAgentå¨æ¯ä¸ªå³çç¹é½è½è¿è¡èªæè¯ä¼°ï¼å¨ä»»å¡ç»æåè¿è¡ç³»ç»æ§å¤çãä¸ãåææºå¶çå®ç°ï¼ä»é误ä¸å¦ä¹3.1 åéåæ模å¼åææºå¶é常å å«ä¸¤ç§æ¨¡å¼ï¼å®æ¶åæï¼In-the-moment Reflectionï¼ï¼å¨æ§è¡è¿ç¨ä¸åç°ä½ç½®ä¿¡åº¦æå¼å¸¸ç»ææ¶ç«å³è§¦åãéç¨äºæ¶é´ææä½éè¦å³æ¶ä¿®æ£çåºæ¯ãäºååæï¼Post-hoc Reflectionï¼ï¼ä»»å¡å®æå对å®æ´è½¨è¿¹è¿è¡ç³»ç»æ§åæãéç¨äºåç°æ·±å±çç¥ç¼ºé·åæç¼éç¨è§å¾ãfrom typing import List, Tuple, Optional from dataclasses import dataclass dataclass class ActionStep: state: dict action: str result: dict confidence: float timestamp: float class Reflector: def __init__(self
调试AI代码？——2024最危险的3个AI Debug陋习，第2个95%工程师每天都在犯（立即停用清单）)
