2026_DesignAI/officefile/md/supplements/03-autonomous-design/03.2-agent-design-patterns.md

03.2 Agent设计模式

核心问题

什么使一个系统成为"Agent"而非简单的程序？ 不同类型的Agent有何区别，各适用于什么场景？ 如何选择合适的Agent架构来解决空间分析问题？

---

概念讲解

Agent的本质

Agent 是能够感知环境并采取行动以实现目标的实体：

┌─────────────────────────────────────────────────────────────┐
│                      Agent的基本结构                         │
├─────────────────────────────────────────────────────────────┤
│                                                             │
│    ┌─────────┐     ┌─────────┐     ┌─────────┐            │
│    │  Sensors │ ──→ │  Agent  │ ──→ │Actuators│            │
│    │  (感知)  │     │ (决策)   │     │ (行动)  │            │
│    └─────────┘     └────┬────┘     └─────────┘            │
│                          │                                  │
│                          ↓                                  │
│                    ┌──────────┐                            │
│                    │Environment│                           │
│                    │   (环境)  │                            │
│                    └──────────┘                            │
│                                                             │
│    感知-决策-行动循环 (Perception-Decision-Action Loop)      │
│                                                             │
└─────────────────────────────────────────────────────────────┘

Agent vs. 程序：

特征	普通程序	Agent
控制流	调用者驱动	自主驱动
状态	被动存储	主动维护世界模型
目标	无明确目标	有内在目标
环境	不感知	持续感知
适应性	固定行为	可学习适应

四种经典Agent类型

根据Russell & Norvig的AI教材，Agent有四种基本设计模式：

Agent类型演进

    Reflex (反应式)
         │
         ├──→ 无状态，直接映射
         │    感知→规则→行动
         │
         ↓
    Model-based (基于模型)
         │
         ├──→ 有内部状态
         │    感知→状态更新→行动
         │
         ↓
    Goal-based (基于目标)
         │
         ├──→ 有目标导向的规划
         │    状态+目标→规划→行动
         │
         ↓
    Utility-based (基于效用)
         │
         └──→ 有量化评估
              状态+目标+效用→最优决策

---

设计原理

1. Reflex Agent（反应式Agent）

特点：直接将感知映射到行动，无内部状态

class ReflexAgent:
    """
    反应式Agent：最简单的Agent类型

    适用场景：
    - 环境完全可观察
    - 当前行动只依赖当前感知
    - 不需要历史信息
    """

    def __init__(self, rules: dict):
        """
        Args:
            rules: {condition: action} 映射规则
        """
        self.rules = rules

    def act(self, percept: dict) -> str:
        """
        根据当前感知选择行动

        Args:
            percept: 当前感知状态

        Returns:
            选择的行动
        """
        for condition, action in self.rules.items():
            if self._match_condition(condition, percept):
                return action

        return self.default_action()

    def _match_condition(self, condition: dict, percept: dict) -> bool:
        """检查条件是否匹配"""
        for key, value in condition.items():
            if percept.get(key) != value:
                return False
        return True

    def default_action(self) -> str:
        """默认行动"""
        return "wait"


# 示例：简单的土地覆盖分类Agent
class LandCoverReflexAgent(ReflexAgent):
    """基于NDVI的土地覆盖分类Agent"""

    def __init__(self):
        rules = {
            {'ndvi_high': True}: 'vegetation',
            {'ndvi_low': True, 'nir_high': True}: 'water',
            {'ndvi_low': True, 'temperature_high': True}: 'urban',
        }
        super().__init__(rules)

    def classify(self, ndvi: float, nir: float, temperature: float) -> str:
        """分类土地覆盖类型"""
        percept = {
            'ndvi_high': ndvi > 0.4,
            'ndvi_low': ndvi <= 0.4,
            'nir_high': nir > 0.3,
            'temperature_high': temperature > 25
        }
        return self.act(percept)

优点：

• 简单高效
• 响应快速
• 易于理解和调试

缺点：

• 无法处理部分可观察环境
• 无法规划未来行动
• 规则冲突时难以决策

2. Model-based Agent（基于模型的Agent）

特点：维护内部状态，跟踪世界的部分不可观察方面

class ModelBasedAgent:
    """
    基于模型的Agent：维护世界状态

    适用场景：
    - 环境部分可观察
    - 需要跟踪历史信息
    - 需要推断隐藏状态
    """

    def __init__(self, transition_model, sensor_model):
        """
        Args:
            transition_model: 状态转移模型 P(s'|s,a)
            sensor_model: 传感器模型 P(o|s)
        """
        self.state = None
        self.transition_model = transition_model
        self.sensor_model = sensor_model
        self.history = []

    def update_state(self, action: str, percept: dict):
        """
        更新内部状态

        使用贝叶斯推断：
        P(s'|o,a,s) ∝ P(o|s') * Σ P(s'|s,a) * P(s)
        """
        if self.state is None:
            # 初始化状态
            self.state = self.sensor_model.estimate(percept)
        else:
            # 预测：基于转移模型
            predicted = self.transition_model.predict(self.state, action)

            # 更新：基于感知
            self.state = self.sensor_model.update(predicted, percept)

        self.history.append({
            'action': action,
            'percept': percept,
            'state': self.state
        })

    def act(self, percept: dict) -> str:
        """选择行动"""
        self.update_state(self.last_action, percept)
        return self._choose_action()

    def _choose_action(self) -> str:
        """基于当前状态选择行动"""
        raise NotImplementedError


# 示例：生态变化检测Agent
class EcologicalChangeAgent(ModelBasedAgent):
    """检测生态系统变化的Agent"""

    class TransitionModel:
        """状态转移模型"""
        def predict(self, state, action):
            # 简单的马尔可夫假设
            new_state = state.copy()
            if action == 'monitor':
                # 状态可能自然变化
                new_state['change_probability'] *= 0.95
            return new_state

    class SensorModel:
        """传感器模型"""
        def estimate(self, percept):
            return {
                'baseline': percept['ndvi'],
                'change_probability': 0.0,
                'confidence': percept['quality']
            }

        def update(self, predicted, percept):
            # 融合预测和观测
            alpha = 0.7  # 预测权重
            new_ndvi = alpha * predicted['baseline'] + (1-alpha) * percept['ndvi']

            change_prob = predicted['change_probability']
            if abs(new_ndvi - predicted['baseline']) > 0.1:
                change_prob += 0.2

            return {
                'baseline': new_ndvi,
                'change_probability': min(1.0, change_prob),
                'confidence': predicted['confidence']
            }

    def __init__(self):
        super().__init__(self.TransitionModel(), self.SensorModel())
        self.last_action = None

    def _choose_action(self) -> str:
        """基于变化概率选择行动"""
        if self.state['change_probability'] > 0.6:
            return 'alert'
        elif self.state['change_probability'] > 0.3:
            return 'investigate'
        else:
            return 'monitor'

3. Goal-based Agent（基于目标的Agent）

特点：显式表示目标，能够规划行动序列

class GoalBasedAgent(ModelBasedAgent):
    """
    基于目标的Agent：有明确的追求目标

    适用场景：
    - 需要规划多步行动
    - 有明确的目标状态
    - 需要考虑行动后果
    """

    def __init__(self, transition_model, sensor_model, planner):
        """
        Args:
            planner: 规划器，搜索从当前状态到目标的路径
        """
        super().__init__(transition_model, sensor_model)
        self.planner = planner
        self.current_goal = None
        self.current_plan = []

    def set_goal(self, goal: dict):
        """设置目标"""
        self.current_goal = goal
        self.current_plan = []
        return self

    def act(self, percept: dict) -> str:
        """选择行动"""
        self.update_state(self.last_action, percept)

        # 检查是否达到目标
        if self._goal_achieved():
            return 'goal_reached'

        # 如果没有计划或计划过时，重新规划
        if not self.current_plan or self._plan_stale():
            self.current_plan = self.planner.plan(
                self.state,
                self.current_goal
            )

        # 执行计划的下一步
        if self.current_plan:
            action = self.current_plan.pop(0)
            self.last_action = action
            return action

        return 'no_plan'

    def _goal_achieved(self) -> bool:
        """检查目标是否达成"""
        if not self.current_goal:
            return False
        return all(
            self.state.get(k) == v
            for k, v in self.current_goal.items()
        )

    def _plan_stale(self) -> bool:
        """检查计划是否需要更新"""
        # 简化版本：检查最近的状态变化
        if len(self.history) < 2:
            return False
        # 实际实现会更复杂
        return False


# 示例：保护区选址Agent
class ReserveSiteAgent(GoalBasedAgent):
    """寻找最佳保护区选址的Agent"""

    class Planner:
        """前向搜索规划器"""
        def plan(self, current_state, goal):
            """
            使用前向搜索规划

            返回行动序列：[action1, action2, ...]
            """
            plan = []

            # 目标：找到至少3个候选地点
            while current_state.get('n_candidates', 0) < goal.get('min_sites', 3):
                # 下一步行动
                if not current_state.get('searched_regions', []):
                    action = ('search_region', 0)
                else:
                    next_region = max(current_state['searched_regions']) + 1
                    action = ('search_region', next_region)

                plan.append(action)
                # 模拟状态更新
                current_state = self._simulate(current_state, action)

                if next_region > 10:  # 防止无限循环
                    break

            return plan

        def _simulate(self, state, action):
            """模拟行动后的状态"""
            new_state = state.copy()
            if action[0] == 'search_region':
                searched = state.get('searched_regions', [])
                searched.append(action[1])
                new_state['searched_regions'] = searched
                # 模拟可能发现候选点
                if action[1] % 3 == 0:  # 每3个区域有1个候选
                    new_state['n_candidates'] = state.get('n_candidates', 0) + 1
            return new_state

    def __init__(self):
        super().__init__(
            super().TransitionModel(),
            super().SensorModel(),
            self.Planner()
        )
        self.last_action = None

    def find_reserve_sites(self, min_sites: int = 3, budget: float = 1000000):
        """寻找保护区选址"""
        self.set_goal({
            'min_sites': min_sites,
            'budget': budget,
            'status': 'found'
        })
        return self

4. Utility-based Agent（基于效用的Agent）

特点：使用效用函数量化目标状态的价值，处理冲突目标

class UtilityBasedAgent(GoalBasedAgent):
    """
    基于效用的Agent：量化目标价值

    适用场景：
    - 有多个冲突目标
    - 目标有不同重要性
    - 需要在不确定环境下决策
    """

    def __init__(self, transition_model, sensor_model, planner, utility_fn):
        """
        Args:
            utility_fn: 效用函数 U(state) → 实数
        """
        super().__init__(transition_model, sensor_model, planner)
        self.utility_fn = utility_fn

    def set_preferences(self, preferences: dict):
        """设置偏好（权重）"""
        self.utility_fn.set_weights(preferences)
        return self

    def evaluate_plan(self, plan: list) -> float:
        """评估计划的期望效用"""
        expected_state = self.state
        total_utility = 0

        for action in plan:
            # 模拟行动
            expected_state = self.transition_model.predict(
                expected_state, action
            )
            # 累积效用
            total_utility += self.utility_fn(expected_state)

        return total_utility

    def choose_best_action(self, available_actions: list) -> str:
        """选择效用最大的行动"""
        best_action = None
        best_utility = float('-inf')

        for action in available_actions:
            # 预测行动后的状态
            predicted_state = self.transition_model.predict(
                self.state, action
            )
            # 计算效用
            utility = self.utility_fn(predicted_state)

            if utility > best_utility:
                best_utility = utility
                best_action = action

        return best_action


class UtilityFunction:
    """效用函数"""

    def __init__(self, objectives: dict):
        """
        Args:
            objectives: {name: (weight, function)}
        """
        self.objectives = objectives

    def set_weights(self, weights: dict):
        """更新目标权重"""
        for name, weight in weights.items():
            if name in self.objectives:
                old_weight, fn = self.objectives[name]
                self.objectives[name] = (weight, fn)

    def __call__(self, state: dict) -> float:
        """计算状态的总效用"""
        total = 0
        for (weight, fn) in self.objectives.values():
            total += weight * fn(state)
        return total


# 示例：土地利用规划Agent
class LandUsePlanningAgent(UtilityBasedAgent):
    """土地利用规划Agent，平衡多个目标"""

    def __init__(self):
        # 定义多个目标
        objectives = {
            'economic': (0.3, self._economic_value),
            'ecological': (0.4, self._ecological_value),
            'social': (0.3, self._social_value),
        }

        super().__init__(
            super().TransitionModel(),
            super().SensorModel(),
            super().Planner(),
            UtilityFunction(objectives)
        )

    @staticmethod
    def _economic_value(state: dict) -> float:
        """经济价值：开发土地产生的收益"""
        return state.get('developed_area', 0) * 1000

    @staticmethod
    def _ecological_value(state: dict) -> float:
        """生态价值：保护的自然栖息地"""
        return -state.get('habitat_loss', 0) * 500

    @staticmethod
    def _social_value(state: dict) -> float:
        """社会价值：住房供应和公共空间"""
        housing = state.get('housing_units', 0)
        green_space = state.get('green_space_ratio', 0)
        return housing * 100 + green_space * 2000

    def plan_land_use(self, area: float, economic_weight: float = 0.3):
        """规划土地利用"""
        self.set_preferences({
            'economic': economic_weight,
            'ecological': 1 - economic_weight - 0.3,
            'social': 0.3
        })
        return self

---

代码示例

完整的四类Agent对比演示

"""
四种Agent类型的完整对比演示

场景：生态监测站需要决定每日行动
"""
import numpy as np
from typing import Dict, List, Optional, Tuple
from dataclasses import dataclass
from enum import Enum


class SensorReading(Enum):
    """传感器读数类型"""
    NORMAL = "normal"
    ANOMALY_DETECTED = "anomaly"
    CRITICAL = "critical"


@dataclass
class EnvironmentState:
    """环境状态"""
    temperature: float
    humidity: float
    species_count: int
    vegetation_health: float  # 0-1
    detected_anomaly: bool
    time_step: int


class MonitoringStation:
    """模拟生态监测站环境"""

    def __init__(self):
        self.state = EnvironmentState(
            temperature=25.0,
            humidity=60.0,
            species_count=15,
            vegetation_health=0.8,
            detected_anomaly=False,
            time_step=0
        )
        self.anomaly_schedule = [5, 12, 18]  # 预定异常发生时间

    def step(self, action: str) -> Tuple[SensorReading, EnvironmentState]:
        """执行一步模拟"""
        self.state.time_step += 1

        # 环境动态变化
        self.state.temperature += np.random.normal(0, 1)
        self.state.humidity += np.random.normal(0, 2)
        self.state.vegetation_health = max(0, min(1,
            self.state.vegetation_health + np.random.normal(0, 0.05)
        ))

        # 检查是否发生异常
        if self.state.time_step in self.anomaly_schedule:
            self.state.detected_anomaly = True
            self.state.vegetation_health -= 0.2

        # 行动影响
        if action == "collect_sample":
            self.state.species_count += np.random.randint(-1, 2)
        elif action == "irrigate":
            self.state.humidity = min(100, self.state.humidity + 10)
            self.state.vegetation_health = min(1, self.state.vegetation_health + 0.05)

        # 生成传感器读数
        reading = self._get_sensor_reading()

        return reading, self.state.copy()

    def _get_sensor_reading(self) -> SensorReading:
        """生成传感器读数"""
        if self.state.vegetation_health < 0.3:
            return SensorReading.CRITICAL
        elif self.state.detected_anomaly:
            return SensorReading.ANOMALY_DETECTED
        return SensorReading.NORMAL

    def reset(self):
        """重置环境"""
        self.__init__()


# ==================== 1. Reflex Agent ====================

class ReflexMonitoringAgent:
    """反应式监测Agent"""

    def __init__(self):
        self.rules = {
            SensorReading.CRITICAL: "emergency_response",
            SensorReading.ANOMALY_DETECTED: "investigate",
            SensorReading.NORMAL: "routine_check"
        }

    def act(self, reading: SensorReading) -> str:
        """根据读数直接行动"""
        return self.rules.get(reading, "routine_check")


# ==================== 2. Model-based Agent ====================

class ModelBasedMonitoringAgent:
    """基于模型的监测Agent"""

    def __init__(self):
        self.belief_state = {
            'anomaly_active': False,
            'anomaly_duration': 0,
            'vegetation_trend': 'stable',
            'last_reading': None
        }

    def act(self, reading: SensorReading) -> str:
        """更新信念状态并行动"""
        # 更新内部状态
        if reading == SensorReading.ANOMALY_DETECTED:
            self.belief_state['anomaly_active'] = True
            self.belief_state['anomaly_duration'] += 1
        elif reading == SensorReading.NORMAL:
            if self.belief_state['anomaly_active']:
                self.belief_state['anomaly_duration'] -= 1
                if self.belief_state['anomaly_duration'] <= 0:
                    self.belief_state['anomaly_active'] = False

        # 基于信念状态决策
        if self.belief_state['anomaly_active']:
            if self.belief_state['anomaly_duration'] > 2:
                return "intensive_monitoring"
            return "investigate"
        return "routine_check"


# ==================== 3. Goal-based Agent ====================

class GoalBasedMonitoringAgent:
    """基于目标的监测Agent"""

    def __init__(self):
        self.current_goal = None
        self.belief_state = {
            'data_coverage': 0.0,
            'samples_collected': 0,
            'anomalies_investigated': 0
        }

    def set_goal(self, goal: str):
        """设置当前目标"""
        self.current_goal = goal
        return self

    def act(self, reading: SensorReading) -> str:
        """基于目标选择行动"""
        # 目标导向的规划
        if self.current_goal == "comprehensive_survey":
            if self.belief_state['data_coverage'] < 1.0:
                return "collect_sample"
            elif self.belief_state['samples_collected'] < 10:
                return "collect_sample"
            else:
                return "compile_report"

        elif self.current_goal == "anomaly_investigation":
            if reading == SensorReading.ANOMALY_DETECTED:
                self.belief_state['anomalies_investigated'] += 1
                return "investigate"
            return "search_for_anomalies"

        return "routine_check"


# ==================== 4. Utility-based Agent ====================

class UtilityMonitoringAgent:
    """基于效用的监测Agent"""

    def __init__(self):
        self.belief_state = {
            'anomaly_active': False,
            'data_coverage': 0.0,
            'resource_remaining': 100,
            'scientific_value': 0
        }
        self.weights = {
            'safety': 0.5,
            'science': 0.3,
            'efficiency': 0.2
        }

    def utility(self, action: str, reading: SensorReading) -> float:
        """计算行动的效用"""
        utility = 0

        # 安全效用
        if reading == SensorReading.CRITICAL:
            if action == "emergency_response":
                utility += 100 * self.weights['safety']
            else:
                utility -= 50 * self.weights['safety']
        elif reading == SensorReading.ANOMALY_DETECTED:
            if action == "investigate":
                utility += 30 * self.weights['safety']
            elif action == "routine_check":
                utility -= 20 * self.weights['safety']

        # 科学价值效用
        if action == "collect_sample":
            if self.belief_state['data_coverage'] < 0.8:
                utility += 20 * self.weights['science']
            else:
                utility += 5 * self.weights['science']

        # 效率效用
        if self.belief_state['resource_remaining'] < 20:
            if action == "routine_check":
                utility += 10 * self.weights['efficiency']
            elif action == "collect_sample":
                utility -= 15 * self.weights['efficiency']

        return utility

    def act(self, reading: SensorReading) -> str:
        """选择效用最大的行动"""
        actions = ["emergency_response", "investigate", "collect_sample",
                   "routine_check", "rest"]

        best_action = "routine_check"
        best_utility = float('-inf')

        for action in actions:
            u = self.utility(action, reading)
            if u > best_utility:
                best_utility = u
                best_action = action

        return best_action


# ==================== 演示对比 ====================

def compare_agents(n_steps: int = 20):
    """对比四种Agent的表现"""
    print("=" * 60)
    print("四种Agent类型在生态监测任务中的对比")
    print("=" * 60)

    agents = {
        "Reflex": ReflexMonitoringAgent(),
        "Model-based": ModelBasedMonitoringAgent(),
        "Goal-based": GoalBasedMonitoringAgent().set_goal("comprehensive_survey"),
        "Utility-based": UtilityMonitoringAgent()
    }

    results = {name: [] for name in agents.keys()}

    for step in range(n_steps):
        env = MonitoringStation()

        for name, agent in agents.items():
            reading, state = env.step("observe")
            action = agent.act(reading)
            results[name].append(action)

    # 打印结果对比
    print("\n行动序列对比:")
    print("-" * 60)
    print(f"{'时间':<6} {'Reflex':<20} {'Model-based':<20}")
    print("-" * 60)

    for i in range(n_steps):
        print(f"{i:<6} {results['Reflex'][i]:<20} {results['Model-based'][i]:<20}")

    print("-" * 60)
    print(f"{'时间':<6} {'Goal-based':<20} {'Utility-based':<20}")
    print("-" * 60)

    for i in range(n_steps):
        print(f"{i:<6} {results['Goal-based'][i]:<20} {results['Utility-based'][i]:<20}")

    # 统计分析
    print("\n行动统计:")
    print("-" * 60)
    for name, actions in results.items():
        from collections import Counter
        counts = Counter(actions)
        print(f"\n{name}:")
        for action, count in counts.most_common():
            print(f"  {action}: {count}")


if __name__ == "__main__":
    compare_agents()

---

案例分析

Claude Code的Agent架构

Claude Code是一个典型的Utility-based Agent，它结合了多种设计模式：

"""
Claude Code的Agent架构简化示意
"""

class ClaudeCodeAgent:
    """
    Claude Code Agent设计

    特点：
    - Model-based: 维护对话上下文状态
    - Goal-based: 追求用户的任务目标
    - Utility-based: 平衡正确性、效率、安全性
    """

    def __init__(self):
        # 内部状态
        self.state = {
            'conversation_history': [],
            'workspace_state': {},  # 文件系统状态
            'tool_results': [],
            'user_goal': None
        }

        # 效用函数组件
        self.utility_components = {
            'task_completion': 0.5,    # 完成任务
            'correctness': 0.3,        # 正确性
            'safety': 0.2,             # 安全性
        }

    def perceive(self, user_input: str, tool_outputs: list):
        """感知：更新内部状态"""
        self.state['conversation_history'].append({
            'role': 'user',
            'content': user_input
        })
        self.state['tool_results'] = tool_outputs

    def plan(self):
        """规划：生成行动计划"""
        # 分析用户意图
        intent = self._analyze_intent()

        # 生成候选行动序列
        candidates = self._generate_candidates(intent)

        # 评估每个候选
        best_plan = max(
            candidates,
            key=lambda p: self._evaluate_plan(p)
        )

        return best_plan

    def act(self, plan: list):
        """执行：按计划调用工具"""
        results = []
        for action in plan:
            result = self._execute_action(action)
            results.append(result)
        return results

    def _evaluate_plan(self, plan: list) -> float:
        """评估计划的效用"""
        utility = 0
        for component, weight in self.utility_components.items():
            if component == 'safety':
                # 检查危险操作
                if any(a.get('dangerous') for a in plan):
                    utility -= 100 * weight
            # ... 其他评估
        return utility

ENAgent的混合设计

ENAgent（生态网络分析Agent）采用了多层混合架构：

class ENAgent:
    """
    ENAgent: 多层混合Agent架构

    反应层：处理简单操作
    规划层：处理复杂分析流程
    效用层：优化分析参数
    """

    def __init__(self):
        # 反应式处理简单命令
        self.reflex_layer = ReflexLayer({
            'load_data': self._load_data,
            'show_status': self._show_status,
        })

        # 规划层处理复杂流程
        self.planning_layer = PlanningLayer()
        self.planning_layer.set_goal('build_ecological_network')

        # 效用层优化参数
        self.utility_layer = UtilityLayer({
            'accuracy': self._accuracy_fn,
            'computation_time': self._time_fn,
            'memory_usage': self._memory_fn,
        })

    def process(self, user_command: str):
        """处理用户命令"""
        # 1. 反应层快速响应
        if user_command in self.reflex_layer.handlers:
            return self.reflex_layer.handle(user_command)

        # 2. 规划层生成流程
        plan = self.planning_layer.generate_plan(user_command)

        # 3. 效用层优化参数
        optimized_plan = self.utility_layer.optimize(plan)

        # 4. 执行计划
        return self._execute(optimized_plan)

---

反思与延伸

思考问题

1. Agent类型选择：你的空间分析项目适合哪种Agent类型？

2. 状态管理：如何在部分可观察环境中维护准确的内部状态？

3. 目标冲突：当生态保护与经济发展冲突时，如何量化权衡？

4. 规划成本：复杂规划的计算成本何时超过了其收益？

延伸阅读

• "Artificial Intelligence: A Modern Approach" (Russell & Norvig) - Chapter 2: Intelligent Agents
• "Agent-Based Modeling" (Railsback & Grimm) - 基于Agent的建模
• ReAct论文 - "ReAct: Synergizing Reasoning and Acting in Language Models"

---

关键要点

1. Agent的核心特征是自主性、感知-行动循环和目标导向
2. Reflex Agent最简单，适合完全可观察环境
3. Model-based Agent通过内部状态处理部分可观察性
4. Goal-based Agent能够规划多步行动达成目标
5. Utility-based Agent通过效用函数处理多目标冲突
6. 实际系统常采用混合架构，结合多种设计模式