[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-18fe92a5-3317-4f4a-ab2b-988c535ccbb9":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":23,"created_at":24,"modified_at":25,"is_published":26,"publish_type":27,"image_url":13,"view_count":28},"18fe92a5-3317-4f4a-ab2b-988c535ccbb9","Google ADK 2.0 把「确定性」焊进 Agent 框架：50% token 省下来的工程范式跃迁","Google 工程师在《Why we built ADK 2.0》中抛出一个反直觉的命题：在企业生产环境中，LLM 不应该是「编排者」，而应该是「工人」。ADK 2.0（Agent Development Kit 2.0）的核心是一套结构化工作流运行时——开发者把「调用第三方 API」这种确定性步骤直接写成代码节点，把「理解客户投诉」这种需要语义推理的任务交给 LLM 节点，再用图把节点按业务逻辑串起来。\n\nGoogle 工程师用「客户退款」做了对比：传统 autonomous agent 把 5 个步骤全塞进一个 LLM 循环，token 用 5,152；改成 ADK 2.0 workflow 后，LLM 只负责「判断政策合规」和「起草邮件」两个语义节点，token 直接压到 2,265，省了约 50%，延迟从 7.2 秒降到 5.7 秒。\n\n更重要的是，这种「确定性骨架 + LLM 软组织」的结构天然解决了三个生产痛点：上下文膨胀（每次 tool 输出污染上下文）、执行脱轨（agent 跳过关键步骤或陷入循环）、提示注入（恶意输入劫持 LLM 走错分支）。Workflow 的边和节点是写死的代码，LLM 哪怕被注入攻击，也调不到 workflow 里没有的边。\n\nADK 2.0 还引入动态工作流：不再强制把复杂逻辑塞进静态路由表，允许直接用 Python asyncio 控制流来表达，再用子 workflow 嵌套。整个框架支持 Python\u002FJava\u002FGo\u002FTypeScript\u002FKotlin 五种语言，Go 版本刚发布。\n\n这套范式的真正价值在于：把「是否要用 Agent」这个二选一，变成了「哪一步用 Agent」。对工程团队来说，可能是 2026 年最有用的认知校正——能用确定性代码就别浪费 token。","https:\u002F\u002Fdevelopers.googleblog.com\u002Fwhy-we-built-adk-20\u002F","3318cb52-f01e-4c9e-a34a-5dbc9fa986f2",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"id":18,"name":19,"slug":19,"description":13,"color":13},"8cf7490f-2449-4ba7-be19-61befa0d92b4","google",{"id":21,"name":22,"slug":22,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm","2026-07-03T02:00:00Z","2026-07-02T18:08:30.925622Z","2026-07-02T18:08:30.925631Z",true,"agent",1]