AI Agents · 课时

研究前沿：AGI 及 beyond

Q: 「研究前沿：AGI 及 beyond」这节课中我会学到什么？

智能体稳健性、长时程记忆和多智能体协作中的开放问题 你通过在浏览器中直接运行的动手代码来练习 AI Agents，全天候 AI 导师会在你学习这节课的过程中回答你的问题。

智能体稳健性、长时程记忆和多智能体协作中的开放问题

研究前沿：AGI 及 beyond 是 CoddyKit 上的免费 AI Agents 课时。这是第 4 节课，共 4 节。你可以在下方免费阅读本课时的完整内容 — 然后在浏览器中使用内置代码编辑器和全天候 AI 导师进行实践。这是 AI Agents 学习路径的一部分，你的进度在网页和 CoddyKit 应用中同步。 AI Agents 课程共包含 4 节课。

2025 年人工智能智能体的现状

截至 2025 年，由大语言模型驱动的人工智能智能体已经能够可靠地完成复杂的多步骤任务、使用工具、进行跨模态推理，并在有限监督下运行。然而，在智能体达到真正的通用能力之前，仍有几个根本性挑战尚未解决。

本课将概览定义下一代人工智能的开放研究前沿。

开放问题一：长时程记忆

当前 LLM 的上下文窗口为 128K–1M 个令牌，虽然令人印象深刻，但对于持续数月的任务来说仍然有限。尚未解决的问题是：如何可靠地对真正的长时程记忆进行压缩、检索和推理，同时避免丢失关键细节或引入幻觉。

# Illustration of long-horizon memory challenges:

LONG_HORIZON_CHALLENGES = {
    'compression': {
        'problem': 'Summarising months of interactions loses nuance',
        'current_approach': 'Hierarchical summarisation (recent detail, old summary)',
        'limitation': 'Important details get compressed away; hallucination risk in summaries'
    },
    'retrieval': {
        'problem': 'Finding the relevant memory among millions of entries',
        'current_approach': 'Embedding-based similarity search (vector databases)',
        'limitation': 'Semantic similarity does not always match relevance; false negatives'
    },
    'reasoning_over_time': {
        'problem': 'Connecting observations from 6 months apart',
        'current_approach': 'Temporal indexing + LLM reasoning',
        'limitation': 'LLMs struggle with precise temporal ordering of distant events'
    }
}

for challenge, details in LONG_HORIZON_CHALLENGES.items():
    print(f'{challenge}: {details["limitation"][:80]}')

开放问题二：跨领域的鲁棒性

当前的智能体很脆弱：为客户支持微调的智能体，可能在新领域（医疗、法律、技术）中的类似任务上失败。真正的鲁棒性意味着，即使面对智能体从未明确训练过的任务和领域，它也能表现良好——这是 AGI 的一项关键要求。

# Measuring domain robustness
import statistics

def measure_domain_robustness(agent_fn, test_suite: dict) -> dict:
    """
    test_suite: {domain: [(input, expected_output)]}
    Returns per-domain accuracy and overall robustness score.
    """
    domain_scores = {}
    for domain, cases in test_suite.items():
        correct = 0
        for inp, expected in cases:
            result = agent_fn(inp)
            # Simplified scoring: check if expected phrase is in result
            if expected.lower() in result.lower():
                correct += 1
        domain_scores[domain] = round(correct / len(cases), 3)

    scores = list(domain_scores.values())
    return {
        'domain_scores': domain_scores,
        'mean_accuracy': round(statistics.mean(scores), 3),
        'min_accuracy': min(scores),  # robustness = performance on worst domain
        'variance': round(statistics.variance(scores), 4)
    }

# High variance = brittle (good at some domains, bad at others)
# Low variance + high mean = robust

if __name__ == '__main__':
    def toy_agent(inp):
        return {
            '2+2': 'The answer is 4',
            'capital of France': 'Paris is the capital'
        }.get(inp, 'I do not know')

    test_suite = {
        'math': [('2+2', '4')],
        'geography': [('capital of France', 'paris')],
    }
    result = measure_domain_robustness(toy_agent, test_suite)
    print('Domain scores:', result['domain_scores'])
    print('Mean accuracy:', result['mean_accuracy'])

开放问题三：多智能体协调

专业化智能体组成的网络可以处理超出任何单个智能体能力范围的任务。但协调这些智能体很困难：它们必须高效沟通，避免重复工作，解决冲突，并在没有中心化瓶颈的情况下共享进展。

import anthropic
import json

client = anthropic.Anthropic(api_key='YOUR_API_KEY')

# Simple task negotiation between two agents
def negotiate_task_division(
    task: str,
    agent1_capabilities: list,
    agent2_capabilities: list
) -> dict:
    prompt = (
        f'Task: {task}\n\n'
        f'Agent A capabilities: {agent1_capabilities}\n'
        f'Agent B capabilities: {agent2_capabilities}\n\n'
        'How should this task be divided between Agent A and Agent B?\n'
        'Minimise handoffs. Assign subtasks to the best-suited agent.\n'
        'Return JSON: {"agent_a_tasks": [str], "agent_b_tasks": [str], '
        '"shared_tasks": [str], "handoffs": int}'
    )
    response = client.messages.create(
        model='claude-opus-4-5', max_tokens=512,
        messages=[{'role': 'user', 'content': prompt}]
    )
    return json.loads(response.content[0].text)

# Open research challenge:
# How do agents coordinate without a central planner
# when each has only partial information?

开放问题四：可解释性

我们目前还无法可靠地解释大型神经网络为什么会做出某个特定决定。可解释性研究旨在识别模型内部的回路、概念和推理模式。缺乏可解释性时，对齐和安全只能完全依赖行为测试，而行为测试无法捕捉所有失败模式。

# Practical interpretability techniques available today:

INTERPRETABILITY_TECHNIQUES = {
    'chain_of_thought': {
        'description': 'Ask model to show reasoning steps',
        'limitation': 'CoT may not reflect true internal computation',
        'example': 'Q: Why did you choose action X? A: Because...'
    },
    'attention_visualisation': {
        'description': 'Show which input tokens the model attended to most',
        'limitation': 'Attention != causation; incomplete explanation',
        'example': 'Highlight most attended tokens in a response'
    },
    'logit_lens': {
        'description': 'Read out predictions at each transformer layer',
        'limitation': 'Requires model internals access (not API-accessible)',
        'example': 'Prediction at layer 12 vs layer 24'
    },
    'activation_patching': {
        'description': 'Intervene on specific neurons to find causal circuits',
        'limitation': 'Research technique, not yet practical in production',
        'example': 'Anthropic mechanistic interpretability research'
    }
}

for technique, info in INTERPRETABILITY_TECHNIQUES.items():
    print(f'{technique}: {info["limitation"][:80]}')

当前状况：模型在 2025 年能做什么

2025 年的前沿模型（GPT-4o、Claude Opus 4、Gemini 1.5 Pro）展现出以下能力：在长上下文中进行多步骤推理，可靠地使用工具，理解视觉和音频，在许多专业基准测试中达到接近人类水平的表现，以及有限但真实的代码生成和调试能力。

CAPABILITY_MAP_2025 = {
    'strengths': [
        'Multi-step reasoning (GSM8K, MATH near human performance)',
        'Code generation (HumanEval >90%)',
        'Instruction following (complex multi-part prompts)',
        'Tool use (reliable function calling)',
        'Vision understanding (OCR, chart analysis, scene description)',
        'Context: 128K-1M tokens',
        'Multi-agent orchestration (AutoGen, CrewAI frameworks)'
    ],
    'limitations': [
        'Long-horizon planning (>20 steps degrades significantly)',
        'Reliable factual grounding without hallucination',
        'Consistent reasoning in out-of-distribution domains',
        'True causal reasoning (vs pattern matching)',
        'Self-knowledge of own uncertainty',
        'Physical world understanding without embodiment'
    ]
}

print('Strengths:', len(CAPABILITY_MAP_2025['strengths']))
print('Active limitations:', len(CAPABILITY_MAP_2025['limitations']))

通往 AGI 的道路：关键研究领域

研究人员普遍认同，一个系统若要达到通用人工智能的标准，就必须具备以下能力：从少量示例中高效学习，跨领域广泛泛化，进行因果而非仅相关性的推理，并在开放式环境中保持稳健运行。

AGI_RESEARCH_AREAS = {
    'sample_efficiency': {
        'question': 'How to learn from 10 examples what LLMs need 10M for?',
        'approaches': ['meta-learning', 'few-shot learning', 'in-context learning']
    },
    'causal_reasoning': {
        'question': 'How to distinguish correlation from causation reliably?',
        'approaches': ['causal graphs', 'do-calculus integration', 'intervention-based training']
    },
    'open_world_operation': {
        'question': 'How to act effectively in environments not seen during training?',
        'approaches': ['world models', 'imagination-based planning', 'transfer learning']
    },
    'recursive_self_improvement': {
        'question': 'Can an agent improve its own architecture safely?',
        'approaches': ['neural architecture search', 'prompt optimisation', 'constrained self-modification']
    }
}

for area, info in AGI_RESEARCH_AREAS.items():
    print(f'{area}: {info["question"][:70]}')

对智能体开发者的实际启示

了解研究前沿有助于您做出更好的工程决策：使用思维链让推理过程可检查；针对新领域设计能够平稳处理失败的智能体；为长时程任务内置人类监督；并在可能时优先选择更简单的架构——系统越简单，失败方式就越可预测。

ENGINEERING_PRINCIPLES_FROM_RESEARCH = {
    'long_horizon_memory': (
        'Use hierarchical summaries + vector retrieval. '
        'Set a hard context age limit and revalidate critical facts. '
        'Never trust old memories without verification.'
    ),
    'domain_robustness': (
        'Evaluate your agent on held-out domains before production. '
        'Monitor domain distribution of production inputs. '
        'Fall back to human when input is out-of-distribution.'
    ),
    'multi_agent': (
        'Minimise inter-agent communication. '
        'Use shared state (not message passing) where possible. '
        'Assign clear non-overlapping scopes to each agent.'
    ),
    'interpretability': (
        'Always request chain-of-thought for high-stakes decisions. '
        'Log all tool calls and intermediate reasoning steps. '
        'Build anomaly detection on the CoT stream, not just final output.'
    )
}

for principle, guidance in ENGINEERING_PRINCIPLES_FROM_RESEARCH.items():
    print(f'{principle}: {guidance[:80]}...')

涌现能力与意外发现

涌现能力是指大型模型在没有接受针对性训练的情况下意外出现的能力。例如：上下文学习、算术和思维链推理。这些能力使能力预测变得困难——下一项突破可能会让所有人都感到意外。

# Historical emergent capability timeline (approximate):
EMERGENCE_TIMELINE = [
    {'year': 2020, 'scale': 'GPT-3 (175B)',
     'emergent': 'Few-shot in-context learning without fine-tuning'},
    {'year': 2022, 'scale': 'PaLM (540B)',
     'emergent': 'Chain-of-thought reasoning with step-by-step prompts'},
    {'year': 2023, 'scale': 'GPT-4',
     'emergent': 'Reliable code generation, bar exam performance'},
    {'year': 2024, 'scale': 'Claude 3 Opus, GPT-4o',
     'emergent': 'Reliable multi-step tool use, vision-language integration'},
    {'year': 2025, 'scale': 'Claude Opus 4, GPT-4o class',
     'emergent': 'Extended multi-agent task delegation, agentic autonomy'}
]

for entry in EMERGENCE_TIMELINE:
    print(f'{entry["year"]} ({entry["scale"]}): {entry["emergent"]}')

print('\nKey insight: capabilities can appear suddenly as scale increases — '
      'current limitations may not be permanent.')

安全研究版图

安全研究与能力研究并行开展。目前的重点领域包括：可扩展监督（如何监督比我们更聪明的智能体）、辩论（两个智能体进行辩论，由人类裁判）、放大（递归地使用人工智能帮助人类评估人工智能），以及可解释性（理解模型内部在做什么）。

SAFETY_RESEARCH_AREAS = {
    'scalable_oversight': (
        'Challenge: how do humans supervise agents that are better than us at the task?\n'
        'Approach: break tasks into verifiable sub-problems humans can check\n'
        'Status: active research at Anthropic, DeepMind, OpenAI'
    ),
    'debate': (
        'Challenge: finding truth when the agent is more capable than the evaluator\n'
        'Approach: two AI agents argue for different answers; human judges quality of argument\n'
        'Status: theoretical framework, limited empirical results'
    ),
    'weak_to_strong_generalization': (
        'Challenge: a weak supervisor training a stronger model\n'
        'Approach: show strong model responses can be elicited by weak supervision\n'
        'Status: OpenAI 2024 paper showed promising early results'
    ),
    'interpretability': (
        'Challenge: understanding neural network internals\n'
        'Approach: mechanistic interp, sparse autoencoders, circuit analysis\n'
        'Status: Anthropic found emotion-like representations in Claude'
    )
}

for area, desc in SAFETY_RESEARCH_AREAS.items():
    print(f'{area}:')
    print(f'  {desc.split(chr(10))[0]}')

作为智能体开发者的未来道路

人工智能智能体领域正在快速发展。能够取得成功的开发者将是那些：紧跟研究进展，在考虑监督和对齐的前提下负责任地构建系统，为优雅降级进行设计，并将智能体视为社会技术系统，而不仅仅是软件。

DEVELOPER_ROADMAP = {
    'immediate': [
        'Master prompt engineering + few-shot design',
        'Build reliable tool-use agents with retry + error handling',
        'Implement proper logging, monitoring, and human oversight',
        'Study agent frameworks: LangChain, AutoGen, CrewAI'
    ],
    'next_6_months': [
        'Build multi-agent systems with clear agent scopes',
        'Implement vector memory + episodic reflection',
        'Contribute to open-source agent tooling',
        'Run proper evals: domain robustness, alignment red-teaming'
    ],
    'long_term': [
        'Follow interpretability research (Anthropic, DeepMind papers)',
        'Engage with alignment research community',
        'Build agents that remain human-overseen as capability grows',
        'Contribute to safety-conscious deployment standards'
    ]
}

for horizon, items in DEVELOPER_ROADMAP.items():
    print(f'{horizon}:')
    for item in items:
        print(f'  - {item}')

知识检查

在大语言模型的语境中，涌现能力一词是什么意思？

回顾：研究前沿、AGI 及未来

恭喜您完成整套人工智能智能体课程！本课的最终要点：

开放问题：长时程记忆、领域鲁棒性、多智能体协调、可解释性
当前优势（2025 年）：工具使用、视觉、推理、1M 个令牌的上下文
通往 AGI 的道路：样本效率、因果推理、开放世界运行
安全研究：可扩展监督、辩论、弱到强泛化、可解释性
您的角色：负责任地构建系统，持续监控，并在每个层级都为人类监督进行设计

感谢您完成人工智能智能体课程体系。现在，您已经具备构建复杂、安全且功能强大的智能体系统所需的能力。

常见问题解答

「研究前沿：AGI 及 beyond」课时是免费的吗？

是的 — 「研究前沿：AGI 及 beyond」的完整文本可在网页上免费阅读。要进行交互式练习（内置代码编辑器和全天候 AI 导师）并解锁 AI Agents 课程的其余内容，请升级到 CoddyKit PRO。 AI Agents 课程共包含 4 节课。

「研究前沿：AGI 及 beyond」这节课中我会学到什么？

智能体稳健性、长时程记忆和多智能体协作中的开放问题你通过在浏览器中直接运行的动手代码来练习 AI Agents，全天候 AI 导师会在你学习这节课的过程中回答你的问题。

学习 AI Agents 需要有经验吗？

无需任何先前经验。CoddyKit 上的 AI Agents 课程适合初学者到高级学习者，你可以从这里开始或从头开始，按照自己的节奏学习。这是第 4 节课，共 4 节。

「研究前沿：AGI 及 beyond」课时需要多长时间？

大多数 CoddyKit 课程大约需要 5–10 分钟。每节课都很精短且互动，所以你能稳步进步，并在网页和应用中从离开的地方继续。

我能在这节 AI Agents 课中编写并运行代码吗？

能。每节 AI Agents 课都包含内置代码编辑器，你可以在浏览器中直接编写并运行真实代码，并获得即时 AI 反馈 — 无需本地设置。

此课程中的所有课时

← 返回 AI Agents