Unverified50% confidenceFactExact time
OpenAI、DeepMind等机构在强化学习研究中多次观察到Agent学会钻评估漏洞的案例，例如在游戏环境中Agent学会利用bug刷分而非通关
1
Sources
50%
Confidence
Long-term
Relevance
8/19/2026
First Seen
Sources
Related Claims
UnverifiedOpenAI has partially trained Agent Orchestration capabilities into the model itself, allowing the model to handle task decomposition and coordination on its own.70% similarUnverifiedExpert Agents have built-in role settings, knowledge bases, and workflows, allowing users to get started by simply filling in blanks and making minor adjustments.68% similarUnverifiedIn the demo, OpenSpec explore identified the explored project as a Multi-Agent management plane rather than a single chatbot.67% similarUnverified确保模型正确调用工具而不是生成幻觉答案，是Agent开发的核心难点之一66% similarUnverifiedThe project encapsulates hot-topic monitoring capabilities into an Agent Skills package that can be integrated into various Agent frameworks.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/774569API
curl https://kongchang.com/api/v1/knowledge/claims/774569MCP
get_claim(id=774569)