LangSmith Fleet 新增可分享技能

🔑 增強重點摘要

• •The 'Skills' feature leverages LangSmith's existing tracing and evaluation infrastructure, allowing developers to version-control agent capabilities alongside their performance metrics.
• •Skills are implemented as modular, reusable prompt-and-tool bundles that can be injected into different agent architectures without requiring a full redeployment of the underlying model.
• •The update includes a new 'Skill Registry' within the LangSmith dashboard, enabling role-based access control (RBAC) to manage which team members can modify or deploy specific agent capabilities.

🛠️ 技術深入

• •Skills are stored as JSON-serialized objects containing system prompts, tool definitions (OpenAPI specs), and associated few-shot examples.
• •Integration utilizes LangChain's 'Runnable' interface, allowing Skills to be composed into existing chains using the pipe (|) operator.
• •Version history for Skills is tracked via Git-like commits within the LangSmith backend, supporting rollback to previous skill iterations.
• •Fleet agents utilize a dynamic lookup mechanism to fetch the latest 'production' tagged version of a Skill at runtime, minimizing latency through edge caching.

🔮 前景展望基於引用來源的 AI 分析

⏳ 時間線