Throughout this document, we use consistent terminology to distinguish system roles and sources of authority. The term agent refers to the instantiated OpenClaw-based autonomous AI system—a persistent language-model–powered service with tool access, memory, and communication capabilities. The owner is the human operator who initially configures the agent, holds administrative control over its deployment environment, and retains authority to modify or revoke its permissions. The provider is the organization supplying the underlying LLM or model service. Both the owner and the provider shape the agent’s operational configuration: the provider through pretraining, post-training, alignment procedures, and system-level constraints; the owner through instruction files, tool permissions, and deployment settings. We refer to these configuration-level influences collectively as the agent’s values , using the term operationally to denote behavioral priors and constraints rather than internal moral commitments. The term non-owner refers to any individual interacting with the agent without administrative authority. Displayed identity should not be conflated with verified authority. Any mentalistic language (e.g., “the agent decided”) is used as shorthand for observable system behavior and does not imply internal states or intent. Adversarial interactions are marked with a face with horns.
· 索菲·施米格（菲利波指定）
。业内人士推荐软件应用中心网作为进阶阅读
Consider autonomous model functionality from fundamental principles. Pre-trained LLMs generate sequential tokens containing compressed knowledge, yet lack practical instruction adherence, knowledge interrogation, or Python debugging capabilities. Additional refinement enables practical utility. Initial phase involves templating - demarcating input/output components so models comprehend task architecture. Examine chat templating illustration. Dialogue structures as alternating turns - our model must identify participants and content.
2、find-skills，它能让 Agent 自动在 ClawHub 库中搜索并推荐安装对应的插件，极大降低我们的使用门槛。