{"id":1185172,"date":"2026-08-20T20:10:16","date_gmt":"2026-08-21T03:10:16","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/?post_type=msr-blog-post&#038;p=1185172"},"modified":"2026-09-03T03:05:14","modified_gmt":"2026-09-03T10:05:14","slug":"agent-lightning-v1-0-%e5%bc%80%e6%ba%90%ef%bc%9a%e9%9d%a2%e5%90%91%e7%9c%9f%e5%ae%9eagent-harness%e7%9a%84%e8%bd%bb%e9%87%8f%e7%ba%a7%e5%bc%ba%e5%8c%96%e5%ad%a6%e4%b9%a0%e6%a1%86%e6%9e%b6","status":"publish","type":"msr-blog-post","link":"https:\/\/www.microsoft.com\/en-us\/research\/articles\/agent-lightning-v1-0-%e5%bc%80%e6%ba%90%ef%bc%9a%e9%9d%a2%e5%90%91%e7%9c%9f%e5%ae%9eagent-harness%e7%9a%84%e8%bd%bb%e9%87%8f%e7%ba%a7%e5%bc%ba%e5%8c%96%e5%ad%a6%e4%b9%a0%e6%a1%86%e6%9e%b6\/","title":{"rendered":"Agent Lightning v1.0 \u5f00\u6e90\uff1a\u9762\u5411\u771f\u5b9eAgent Harness\u7684\u8f7b\u91cf\u7ea7\u5f3a\u5316\u5b66\u4e60\u6846\u67b6"},"content":{"rendered":"\n<p class=\"wp-block-paragraph\">\u968f\u7740 AI Agent \u4ece\u5355\u7eaf\u7684\u6a21\u578b\u9010\u6e10\u6f14\u5316\u4e3a\u7531\u6a21\u578b\u3001\u5de5\u5177\u548c\u6267\u884c\u73af\u5883\u5171\u540c\u7ec4\u6210\u7684\u590d\u6742\u7cfb\u7edf\uff0cAgent \u7684\u80fd\u529b\u8d8a\u6765\u8d8a\u4f9d\u8d56\u4e8e\u6a21\u578b\u4e4b\u5916\u7684&nbsp;Agent&nbsp;Harness\uff08Agent\u8fd0\u884c\u6846\u67b6\uff09\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7136\u800c\uff0c\u5f53\u6211\u4eec\u5e0c\u671b\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u8fdb\u4e00\u6b65\u63d0\u5347 Agent \u80fd\u529b\u65f6\uff0c\u4e00\u4e2a\u73b0\u5b9e\u95ee\u9898\u51fa\u73b0\u4e86\uff1a\u8bb8\u591a\u73b0\u6709 Agent RL \u7cfb\u7edf\u8981\u6c42\u5f00\u53d1\u8005\u5c06 Agent \u91cd\u65b0\u5b9e\u73b0\u5230\u8bad\u7ec3\u6846\u67b6\u5185\u90e8\u3002\u8fd9\u4e0d\u4ec5\u589e\u52a0\u4e86\u5de5\u7a0b\u6210\u672c\uff0c\u4e5f\u8ba9\u8bad\u7ec3\u73af\u5883\u4e0e\u771f\u5b9e\u90e8\u7f72\u73af\u5883\u9010\u6e10\u8131\u8282\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86&nbsp;Harnessed Agentic RL&nbsp;\u8bad\u7ec3\u8303\u5f0f\uff0c\u5e76\u5f00\u6e90\u4e86\u5168\u9762\u91cd\u6784\u7684 Agent Lightning v1.0&nbsp;\u3002\u76f8\u6bd4\u539f\u7248\u672c\u7684<a href=\"https:\/\/www.microsoft.com\/en-us\/research\/articles\/agent-lightning\/\" target=\"_blank\" rel=\"noreferrer noopener\">Agent&nbsp;Lightning<\/a>\uff0cv1.0&nbsp;\u66f4\u5f3a\u8c03\u8f7b\u91cf\u3001\u771f\u5b9e\u7684&nbsp;Harness&nbsp;\u96c6\u6210\u4ee5\u53ca\u5b8c\u6574\u53ef\u590d\u73b0\u7684&nbsp;Agent RL&nbsp;\u8bad\u7ec3\u6d41\u7a0b\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c\u5728&nbsp;Agent Lightning v1.0&nbsp;\u4e2d\uff0c\u90e8\u7f72\u65f6\u4f7f\u7528\u4ec0\u4e48 Agent Harness\uff0c\u8bad\u7ec3\u65f6\u5c31\u76f4\u63a5\u8ba9\u8fd9\u4e2a Harness \u53c2\u4e0e\u5f3a\u5316\u5b66\u4e60\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u4e86\u8ba9 Agent RL \u66f4\u8d34\u8fd1\u771f\u5b9e\u4e16\u754c\u4e2d\u7684 Agent \u7cfb\u7edf\uff0cAgent Lightning v1.0 \u56f4\u7ed5&nbsp;Harnessed Agentic RL \u8fdb\u884c\u4e86\u5168\u9762\u91cd\u6784\uff0c\u5728\u7cfb\u7edf\u8bbe\u8ba1\u3001\u5de5\u7a0b\u5b9e\u8df5\u548c\u8bad\u7ec3\u6548\u679c\u7b49\u65b9\u9762\u90fd\u5e26\u6765\u4e86\u591a\u9879\u5173\u952e\u6539\u8fdb\u3002&nbsp;<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6781\u81f4\u8f7b\u91cf<\/strong><strong>\uff1a<\/strong><strong>\u5b8c\u6574\u6846\u67b6<\/strong><strong>\u4ec5<\/strong><strong>\u7ea6 3,500&nbsp;<\/strong><strong>\u884c\u4ee3\u7801<\/strong><strong>\u3002<\/strong>Agent Lightning v1.0&nbsp;\u5c06\u7b80\u5355\u6027\u4f5c\u4e3a\u7b2c\u4e00\u8bbe\u8ba1\u539f\u5219\uff0c\u4ee5\u5c3d\u53ef\u80fd\u5c0f\u4e14\u6e05\u6670\u7684\u4ee3\u7801\u5e93\u5b9e\u73b0\u5b8c\u6574\u7684 Harnessed Agentic RL \u7cfb\u7edf\uff0c\u8ba9\u6846\u67b6\u66f4\u5bb9\u6613\u7406\u89e3\u3001\u4fee\u6539\u548c\u6269\u5c55\u3002&nbsp;<\/li>\n\n\n\n<li><strong>\u76f4\u63a5\u4f7f\u7528\u771f\u5b9e Agent Harness \u8fdb\u884c<\/strong><strong>\u8bad<\/strong><strong>\u7ec3<\/strong><strong>\u3002<\/strong>Agent&nbsp;\u53ef\u4ee5\u901a\u8fc7&nbsp;Agent Lightning v1.0&nbsp;\u63d0\u4f9b\u7684&nbsp;LLM Proxy&nbsp;\u4e0e\u6a21\u578b\u4ea4\u4e92\uff0c\u65e0\u9700\u4fee\u6539\u539f\u6709&nbsp;Agent&nbsp;Harness&nbsp;\u4ee3\u7801\u3002&nbsp;<\/li>\n\n\n\n<li><strong>\u539f\u751f\u652f\u6301 Kubernetes\u3002<\/strong>Agent&nbsp;\u53ef\u4ee5\u76f4\u63a5\u4ee5&nbsp;Kubernetes&nbsp;\u4efb\u52a1\u7684\u5f62\u5f0f\u8fd0\u884c\uff0c\u65e0\u9700\u4f9d\u8d56\u5916\u90e8\u5546\u4e1a\u6c99\u76d2\u670d\u52a1\u3002\u65e0\u8bba\u662f\u81ea\u5efa\u96c6\u7fa4\u8fd8\u662f\u672c\u5730\u57fa\u7840\u8bbe\u65bd\uff0c\u90fd\u80fd\u591f\u652f\u6301\u5927\u89c4\u6a21 Agent rollout\u3002&nbsp;<\/li>\n\n\n\n<li><strong>\u63d0\u4f9b\u5b8c\u6574\u7684 Coding Agent \u8bad\u7ec3\u793a\u4f8b\u3002<\/strong>\u7814\u7a76\u5458\u4eec\u57fa\u4e8e Qwen3.5-9B \u6784\u5efa\u4e86\u4e00\u5957\u7aef\u5230\u7aef\u8bad\u7ec3\u6d41\u7a0b\uff0c\u4ec5\u4f7f\u7528\u7ea6 6,000&nbsp;\u6761\u8bad\u7ec3\u6837\u672c\uff0c\u5c31\u5c06&nbsp;SWE-bench Verified \u4e0a\u7684Pass@1\u4ece 41.8%&nbsp;\u63d0\u5347\u81f3 56.4%\uff0c\u7edd\u5bf9\u63d0\u5347 14.6&nbsp;\u4e2a\u767e\u5206\u70b9\u3002\u540c\u65f6\uff0c\u6570\u636e\u6e05\u6d17\u3001Reward Hacking \u9632\u62a4\u4ee5\u53ca\u8bad\u7ec3\u811a\u672c\u5747\u5df2\u5f00\u6e90\uff0c\u65b9\u4fbf\u793e\u533a\u590d\u73b0\u548c\u8fdb\u4e00\u6b65\u7814\u7a76\u3002&nbsp;<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong><em>\u8bba\u6587\u4fe1\u606f\u5df2\u6574\u7406\u81f3\u6587\u672b\uff0c\u6b22\u8fce\u70b9\u51fb\u76f8\u5173\u94fe\u63a5\uff0c\u4e86\u89e3\u66f4\u591a\u6280\u672f\u8be6\u60c5\u3002<\/em><\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n\n\n\n<h2 id=\"\u5f53-agent-\u8d8a\u6765\u8d8a\u590d\u6742-\u4f20\u7edf-agentic-rl-\u4e3a\u4f55\u5f00\u59cb\u5931\u6548\" class=\"wp-block-heading\">\u5f53 Agent \u8d8a\u6765\u8d8a\u590d\u6742\uff0c\u4f20\u7edf Agentic RL \u4e3a\u4f55\u5f00\u59cb\u5931\u6548\uff1f<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f20\u7edf Agentic RL \u901a\u5e38\u5047\u8bbe\u8bad\u7ec3\u6846\u67b6\u81ea\u5df1\u62e5\u6709\u5b8c\u6574\u7684\u73af\u5883\u4ea4\u4e92\u5faa\u73af\u3002\u4f8b\u5982\uff0c\u5728\u5178\u578b\u7684&nbsp;ReAct&nbsp;Agent&nbsp;\u4e2d\uff0c\u6a21\u578b\u751f\u6210\u52a8\u4f5c\uff08action\uff09\uff0c\u73af\u5883\u8fd4\u56de\u89c2\u5bdf\u7ed3\u679c\uff08observation\uff09\uff0c\u89c2\u5bdf\u7ed3\u679c\u88ab\u8ffd\u52a0\u5230\u5df2\u6709\u4e0a\u4e0b\u6587\u4e2d\uff0c\u968f\u540e\u6a21\u578b\u7ee7\u7eed\u751f\u6210\u4e0b\u4e00\u6b65\u52a8\u4f5c\u3002\u6574\u4e2a rollout&nbsp;\u56e0\u6b64\u5929\u7136\u5bf9\u5e94\u4e00\u6761\u8fde\u7eed\u7684 token&nbsp;\u8f68\u8ff9\u3002\u65e9\u671f\u7684&nbsp;verl\u3001AReaL\u3001slime&nbsp;\u7b49\u5f3a\u5316\u5b66\u4e60\u7cfb\u7edf\u57fa\u672c\u6cbf\u7528\u4e86\u8fd9\u4e00\u8bbe\u8ba1\u3002\u5982\u679c\u60f3\u8bad\u7ec3\u4e00\u4e2a Agent\uff0c\u5f00\u53d1\u8005\u901a\u5e38\u9700\u8981\u628a Agent\u5faa\u73af\u91cd\u65b0\u5b9e\u73b0\u5230 RL \u6846\u67b6\u5185\u90e8\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f46\u73b0\u5b9e\u4e2d\u7684 Agent Harness \u5df2\u7ecf\u8d8a\u6765\u8d8a\u590d\u6742\u3002\u65e0\u8bba\u662fmini-SWE-agent\u3001OpenHands\u3001OpenCode\u3001ClaudeCode\u3001Codex&nbsp;\u7b49\u4ee3\u7801&nbsp;Agent \uff0c\u8fd8\u662f\u5404\u79cd\u901a\u7528 &nbsp;Agent&nbsp;\u7cfb\u7edf\uff0c\u90fd\u62e5\u6709\u72ec\u7acb\u7684\u4e0a\u4e0b\u6587\u7ba1\u7406\u3001\u5de5\u5177\u534f\u8bae\u3001\u6267\u884c\u903b\u8f91\u548c\u8f6f\u4ef6\u4f9d\u8d56\u3002\u4e3a\u4e86 RL \u8bad\u7ec3\u800c\u91cd\u65b0\u5b9e\u73b0\u4e00\u5957 Harness\uff0c\u4e0d\u4ec5\u5de5\u7a0b\u4ee3\u4ef7\u9ad8\u6602\uff0c\u66f4\u91cd\u8981\u7684\u662f\uff0c\u8bad\u7ec3\u65f6 Agent \u7684\u6267\u884c\u65b9\u5f0f\uff0c\u5f88\u53ef\u80fd\u5df2\u7ecf\u4e0e\u771f\u5b9e\u90e8\u7f72\u73af\u5883\u4e2d\u7684\u6267\u884c\u65b9\u5f0f\u4e0d\u540c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u65e9\u671f\u63a2\u7d22\u4e2d\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u89e3\u8026\u601d\u8def\uff1a\u4e0d\u8981\u6c42 Agent \u8fdb\u5165 RL \u6846\u67b6\uff0c\u800c\u662f\u5728 Agent \u4e0e LLMs&nbsp;\u4e4b\u95f4\u63d2\u5165\u4e00\u4e2a\u5927\u6a21\u578b\u7684\u63a5\u53e3\u5730\u5740\uff08LLMs&nbsp;endpoint Proxy\uff09\u3002Agent \u7ee7\u7eed\u6309\u7167\u539f\u6765\u7684\u65b9\u5f0f\u8fd0\u884c\uff0c\u53ea\u9700\u5c06\u539f\u672c\u8c03\u7528\u6a21\u578b API \u7684\u7aef\u70b9\uff08endpoint\uff09\u6307\u5411 Agent Lightning\uff0c\u8bad\u7ec3\u6846\u67b6\u4fbf\u53ef\u4ee5\u89c2\u5bdf\u5e76\u8bb0\u5f55\u6a21\u578b\u8c03\u7528\u3002\u5728 v1.0 \u4e2d\uff0c\u7814\u7a76\u5458\u4eec\u8fdb\u4e00\u6b65\u5c06\u8fd9\u4e00\u8303\u5f0f\u660e\u786e\u5b9a\u4e49\u4e3a&nbsp;Harnessed Agentic RL\uff0c\u5373\u90e8\u7f72\u65f6\u4f7f\u7528\u4ec0\u4e48 Agent Harness\uff0c\u8bad\u7ec3\u65f6\u5c31\u76f4\u63a5\u8ba9\u8fd9\u4e2a Harness \u53c2\u4e0e\u5f3a\u5316\u5b66\u4e60\u3002&nbsp;<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"732\" height=\"330\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/10-1.png\" alt=\"diagram, text, application\" class=\"wp-image-1185173\" style=\"width:770px\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/10-1.png 732w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/10-1-300x135.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/10-1-240x108.png 240w\" sizes=\"auto, (max-width: 732px) 100vw, 732px\" \/><figcaption class=\"wp-element-caption\">\u56fe1\uff1a\u4f20\u7edfAgentic&nbsp;RL\u548cHarnessed Agentic RL\u7684\u5bf9\u6bd4\u3002\u4f20\u7edfAgentic RL\u7531\u8bad\u7ec3\u6846\u67b6\u76f4\u63a5\u7ba1\u7406\u73af\u5883\u548cAgent\u5faa\u73af\uff0c\u800cHarnessed Agentic RL\u7531Harness\u7ba1\u7406\u73af\u5883\u548c\u5faa\u73af\u3002<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n\n\n\n<h2 id=\"\u8ba9\u771f\u5b9e-agent-\u53c2\u4e0e\u8bad\u7ec3-\u5e76\u6ca1\u6709\u60f3\u8c61\u4e2d\u7b80\u5355\" class=\"wp-block-heading\">\u8ba9\u771f\u5b9e Agent \u53c2\u4e0e\u8bad\u7ec3\uff0c\u5e76\u6ca1\u6709\u60f3\u8c61\u4e2d\u7b80\u5355&nbsp;<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Harnessed Agentic RL&nbsp;\u4e0e\u4f20\u7edf&nbsp;Agentic RL&nbsp;\u7684\u4e00\u4e2a\u6838\u5fc3\u533a\u522b\u662f\uff1a\u73af\u5883\u4ea4\u4e92\u5faa\u73af\u7531&nbsp;Agent Harness&nbsp;\u800c\u4e0d\u662f\u8bad\u7ec3\u6846\u67b6\u8d1f\u8d23\u3002\u8bad\u7ec3\u7cfb\u7edf\u53ea\u80fd\u89c2\u5bdf\u4e00\u7cfb\u5217&nbsp;LLMs&nbsp;\u7684\u8bf7\u6c42\u4e0e\u54cd\u5e94\u5bf9\uff0c\u56e0\u6b64\u4e00\u4e2a rollout \u53ef\u80fd\u88ab\u62c6\u6210\u52a8\u6001\u6570\u91cf\u7684\u8bad\u7ec3\u6837\u672c\u3002\u8fd9\u4e5f\u5e26\u6765\u4e86\u56db\u4e2a\u5173\u952e\u6311\u6218\u3002&nbsp;<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u4e0a\u4e0b\u6587\u91cd<\/strong><strong>\u65b0<\/strong><strong>\u5206\u8bcd\u4e0e\u6837\u672c\u5408\u5e76<\/strong><strong>\uff08<\/strong><strong>Retokenization & Sample Merging<\/strong><strong>\uff09<\/strong><strong>\uff1a<\/strong>Harness&nbsp;\u901a\u5e38\u4ee5\u6587\u672c\u5f62\u5f0f\u7ef4\u62a4\u4e0a\u4e0b\u6587\uff0c\u4f46&nbsp;RL&nbsp;\u8bad\u7ec3\u4f9d\u8d56&nbsp;rollout&nbsp;\u65f6\u771f\u5b9e\u91c7\u6837\u7684&nbsp;token&nbsp;IDs\u3002\u5373\u4f7f\u4e24\u6b21\u8c03\u7528\u5728\u6587\u672c\u5c42\u9762\u53ef\u4ee5\u8fde\u7eed\u62fc\u63a5\uff0c\u91cd\u65b0\u7ecf\u8fc7\u5bf9\u8bdd\u6a21\u677f\u548c\u5206\u8bcd\u5668\uff08tokenizer\uff09\u540e\uff0ctoken&nbsp;\u8fb9\u754c\u4e5f\u53ef\u80fd\u53d1\u751f\u53d8\u5316\uff0c\u56e0\u6b64\u76f8\u90bb\u8c03\u7528\u5e76\u4e0d\u603b\u80fd\u5b89\u5168\u5408\u5e76\u6210\u540c\u4e00\u4e2a\u8bad\u7ec3\u6837\u672c\u3002&nbsp;<\/li>\n\n\n\n<li><strong>\u4f18\u52bf\u503c\u8ba1\u7b97\uff08<\/strong><strong>Advantage Calculation<\/strong><strong>\uff09<\/strong><strong>\uff1a<\/strong>\u4e00\u4e2a rollout \u53ef\u80fd\u7531\u4e8e\u91cd\u65b0\u5206\u8bcd\uff08retokenization\uff09\u3001\u5b50\u667a\u80fd\u4f53\uff08subagent\uff09\u6216\u4e0a\u4e0b\u6587\u603b\u7ed3\uff08context summarization\uff09\u88ab\u62c6\u6210\u591a\u4e2a\u6837\u672c\u3002\u5982\u679c\u76f4\u63a5\u5728\u6837\u672c\u5c42\u7ea7\u8ba1\u7b97\u57fa\u7ebf\u548c\u4f18\u52bf\u503c\uff0c\u5c31\u4f1a\u8ba9\u4ea7\u751f\u66f4\u591a\u6837\u672c\u7684 rollout \u88ab\u91cd\u590d\u8ba1\u5165\uff0c\u4ece\u800c\u6539\u53d8\u539f\u672crollout\u5c42\u7ea7\u7684\u7edf\u8ba1\u5173\u7cfb\u3002&nbsp;<\/li>\n\n\n\n<li><strong>\u635f\u5931\u5f52\u4e00\u5316\uff08<\/strong><strong>Loss Normalization<\/strong><strong>\uff09<\/strong><strong>\uff1a<\/strong>\u7c7b\u4f3c\u5730\uff0c\u5982\u679c\u76f4\u63a5\u6309\u7167\u6837\u672c\u6570\u91cf\u5e73\u5747\u635f\u5931\uff0c\u4ea7\u751f\u66f4\u591a\u6837\u672c\u7684 rollout \u4f1a\u83b7\u5f97\u66f4\u5927\u7684\u4f18\u5316\u6743\u91cd\u3002\u7531\u4e8e\u6837\u672c\u6570\u91cf\u5f80\u5f80\u53ea\u662f Harness \u5185\u90e8\u884c\u4e3a\u5e26\u6765\u7684\u7ed3\u679c\uff0c\u56e0\u6b64\u635f\u5931\u5f52\u4e00\u5316\u4e5f\u9700\u8981\u907f\u514d\u88ab\u6837\u672c\u6570\u91cf\u626d\u66f2\u3002&nbsp;<\/li>\n\n\n\n<li><strong>\u8bad\u7ec3<\/strong><strong>\u8d44\u6e90<\/strong><strong>\u8c03\u5ea6\uff08<\/strong><strong>Training Backend Scheduling<\/strong><strong>\uff09<\/strong><strong>\uff1a<\/strong>\u6bcf\u4e2a rollout \u6700\u7ec8\u4f1a\u4ea7\u751f\u591a\u5c11&nbsp;\u6837\u672c\u3001\u6bcf\u4e2a\u6837\u672c\u6709\u591a\u957f\uff0c\u53ea\u6709 Harness \u6267\u884c\u5b8c\u6210\u540e\u624d\u80fd\u786e\u5b9a\u3002\u4f46 GPU&nbsp;\u6570\u91cf\u4ee5\u53ca\u6570\u636e\/\u5f20\u91cf\u5e76\u884c\uff08data \/ tensor parallel\uff09\u914d\u7f6e\u901a\u5e38\u662f\u56fa\u5b9a\u7684\uff0c\u56e0\u6b64\u540e\u7aef\u9700\u8981\u52a8\u6001\u5730\u5c06\u8fd9\u4e9b\u53ef\u53d8\u957f\u5ea6\u3001\u53ef\u53d8\u6570\u91cf\u7684\u6837\u672c\u6620\u5c04\u5230\u56fa\u5b9a\u7684\u8bad\u7ec3\u8d44\u6e90\u4e0a\u3002&nbsp;<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u9488\u5bf9\u8fd9\u4e9b\u95ee\u9898\uff0cAgent Lightning v1.0\u91cd\u65b0\u8bbe\u8ba1\u4e86\u6837\u672c\u6784\u5efa\uff08sample construction\uff09\u3001\u4f18\u52bf\u503c\u3001\u635f\u5931\u5f52\u4e00\u5316\u548c\u8bad\u7ec3\u8c03\u5ea6\uff0c\u5e76\u5c06&nbsp;rollout&nbsp;\u5c42\u7ea7\u4f5c\u4e3a\u91cd\u8981\u7684\u7edf\u8ba1\u548c\u4f18\u5316\u5355\u4f4d\u3002&nbsp;<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n\n\n\n<h2 id=\"\u75283-500\u884c\u4ee3\u7801-\u642d\u5efa\u5b8c\u6574-agent-rl-\u63a7\u5236\u5e73\u9762\" class=\"wp-block-heading\">\u75283,500\u884c\u4ee3\u7801\uff0c\u642d\u5efa\u5b8c\u6574 Agent RL \u63a7\u5236\u5e73\u9762&nbsp;&nbsp;<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u7cfb\u7edf\u8bbe\u8ba1\u4e0a\uff0cAgent Lightning v1.0 \u5c06\u7b80\u5355\u6027\u4f5c\u4e3a\u7b2c\u4e00\u539f\u5219\u3002\u6574\u4e2a\u6846\u67b6\u7ea6 3,500&nbsp;\u884c\u4ee3\u7801\uff0c\u6838\u5fc3\u7531\u4e09\u4e2a\u7ec4\u4ef6\u7ec4\u6210\uff1aAPI&nbsp;\u7f51\u5173\uff08API&nbsp;Gateway\uff09\u3001\u91c7\u6837\u8f68\u8ff9\u63a7\u5236\u5668\uff08Rollout Controller\uff09\u548c\u81ea\u5b9a\u4e49\u8bad\u7ec3\u5668\uff08Customized Trainer\uff09\u3002&nbsp;<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">API Gateway&nbsp;\u8d1f\u8d23\u4fdd\u5b58&nbsp;rollout\u3001\u6a21\u578b\u548c\u4e8b\u4ef6\uff08event\uff09\uff0c\u5e76\u63d0\u4f9b\u517c\u5bb9 OpenAI \u7684\u5927\u6a21\u578b\u4ee3\u7406\u670d\u52a1\u3002Agent Harness&nbsp;\u7684\u6bcf\u6b21\u6a21\u578b\u8c03\u7528\u90fd\u4f1a\u81ea\u52a8\u5173\u8054\u5230\u5bf9\u5e94\u7684rollout\uff0c\u5e76\u8bb0\u5f55\u8bad\u7ec3\u6240\u9700\u7684\u63d0\u793a\u8bcd\u3001\u56de\u590d\u548c\u5bf9\u6570\u6982\u7387\u3002&nbsp;<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Rollout Controller&nbsp;\u8d1f\u8d23\u542f\u52a8\u548c\u7ba1\u7406&nbsp;Agent&nbsp;\u6267\u884c\uff0c\u65e2\u652f\u6301\u672c\u5730\u8fdb\u7a0b\uff0c\u4e5f\u53ef\u4ee5\u76f4\u63a5\u5c06 Agent \u4f5c\u4e3a\u6807\u51c6 Kubernetes&nbsp;\u4efb\u52a1\u8fd0\u884c\uff0c\u4ece\u800c\u5c06 Agent \u6267\u884c\u4e0e RL&nbsp;\u8bad\u7ec3\u5668\u5b8c\u5168\u89e3\u8026\u3002&nbsp;<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Customized Trainer&nbsp;\u57fa\u4e8e&nbsp;VERL&nbsp;\u5b9e\u73b0\uff0c\u8d1f\u8d23\u521b\u5efa&nbsp;rollout\u3001\u7b49\u5f85\u6267\u884c\u5b8c\u6210\u3001\u6536\u96c6\u6837\u672c\uff0c\u5e76\u901a\u8fc7\u6837\u672c\u9002\u914d\u5668\uff08sample&nbsp;adapter\uff09\u6784\u9020\u6700\u7ec8\u7684 RL&nbsp;\u8bad\u7ec3\u6837\u672c\u3002&nbsp;<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u6b64\uff0c\u5bf9\u4e8e\u5df2\u6709 Agent Harness\uff0c\u901a\u5e38\u53ea\u9700\u8981\u5c06\u5927\u6a21\u578b\u7684\u63a5\u53e3\u5730\u5740\u6307\u5411 Agent Lightning \u7684 Proxy\uff0c\u5c31\u53ef\u4ee5\u5feb\u901f\u63a5\u5165 RL \u8bad\u7ec3\u3002&nbsp;<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"865\" height=\"293\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/10-2.png\" alt=\"diagram\" class=\"wp-image-1185174\" style=\"width:770px\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/10-2.png 865w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/10-2-300x102.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/10-2-768x260.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/10-2-240x81.png 240w\" sizes=\"auto, (max-width: 865px) 100vw, 865px\" \/><figcaption class=\"wp-element-caption\">\u56fe2\uff1aAgent&nbsp;Lightning v1.0\u7cfb\u7edf\u67b6\u6784&nbsp;<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n\n\n\n<h2 id=\"gpu-\u5229\u7528\u7387\u7ffb\u500d-collocated-async-rl\" class=\"wp-block-heading\">GPU \u5229\u7528\u7387\u7ffb\u500d\uff1aCollocated&nbsp;Async RL&nbsp;&nbsp;<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Agent&nbsp;\u5de5\u4f5c\u8d1f\u8f7d\u7684\u53e6\u4e00\u4e2a\u7279\u70b9\u662f rollout \u65f6\u95f4\u9ad8\u5ea6\u4e0d\u5747\u5300\u3002\u540c\u6b65 RL \u5fc5\u987b\u7b49\u5f85\u4e00\u4e2a\u6279\u6b21\u4e2d\u6700\u6162\u7684 Agent \u5b8c\u6210\uff0c\u5bb9\u6613\u9020\u6210\u5927\u91cf GPU \u7a7a\u95f2\uff1b\u5b8c\u5168\u5f02\u6b65&nbsp;RL \u867d\u7136\u63d0\u9ad8\u5229\u7528\u7387\uff0c\u5374\u5f80\u5f80\u9700\u8981\u5206\u522b\u7ef4\u62a4 rollout GPU&nbsp;\u6c60\u4e0e\u8bad\u7ec3&nbsp;GPU&nbsp;\u6c60\uff0c\u6240\u9700\u8d44\u6e90\u66f4\u591a\u3002&nbsp;<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5bf9\u6b64\uff0cAgent Lightning v1.0&nbsp;\u63d0\u51fa\u4e86&nbsp;Collocated Async RL\uff08\u5171\u4f4d\u7f6e\u5f02\u6b65\u5f3a\u5316\u5b66\u4e60\uff09\uff0c\u8ba9&nbsp;rollout \u4e0e\u6a21\u578b\u66f4\u65b0\u5171\u4eab\u540c\u4e00\u7ec4 GPU\u3002\u7cfb\u7edf\u6536\u96c6\u5230\u8db3\u591f rollout \u540e\u5f00\u59cb\u66f4\u65b0\uff0c\u6b64\u65f6 API Gateway \u6682\u505c\u63a5\u53d7\u65b0\u7684\u8bf7\u6c42\uff0c\u7b49\u5f85\u6b63\u5728\u6267\u884c\u7684\u8bf7\u6c42\u7ed3\u675f\uff1b\u66f4\u65b0\u5b8c\u6210\u4e4b\u540e\u518d\u6062\u590d rollout\u3002\u6574\u4e2a\u72b6\u6001\u5207\u6362\u8fc7\u7a0b\u5bf9\u5916\u90e8 Agent Harness \u5b8c\u5168\u900f\u660e\u3002&nbsp;<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b9e\u9a8c\u4e2d\uff0c\u8be5\u65b9\u6cd5\u76f8\u6bd4\u540c\u6b65&nbsp;RL \u83b7\u5f97\u4e86\u7ea62\u500d\u7684\u7aef\u5230\u7aef\u52a0\u901f\uff0c\u540c\u65f6\u53c8\u6bd4\u4f20\u7edf\u5f02\u6b65&nbsp;RL \u4f7f\u7528\u4e86\u66f4\u5c11\u7684 GPU\u3002&nbsp;<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"827\" height=\"275\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/10-3.jpg\" alt=\"graphical user interface, chart\" class=\"wp-image-1185175\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/10-3.jpg 827w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/10-3-300x100.jpg 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/10-3-768x255.jpg 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/10-3-240x80.jpg 240w\" sizes=\"auto, (max-width: 827px) 100vw, 827px\" \/><figcaption class=\"wp-element-caption\">\u56fe3\uff1a\u540c\u6b65RL\uff0c\u5f02\u6b65RL\u548cCollocated Async RL\u7684\u5bf9\u6bd4\u3002Collocated Async RL\u5728\u63d0\u9ad8\u8d44\u6e90\u5229\u7528\u7387\u7684\u540c\u65f6\uff0c\u4e5f\u51cf\u5c11\u4e86GPU\u8d44\u6e90\u7684\u5360\u7528\u3002&nbsp;<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n\n\n\n<h2 id=\"\u544a\u522b\u6602\u8d35-sandbox-\u539f\u751f\u652f\u6301-kubernetes\" class=\"wp-block-heading\">\u544a\u522b\u6602\u8d35&nbsp;sandbox\uff1a\u539f\u751f\u652f\u6301 Kubernetes&nbsp;&nbsp;<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 RL \u8bad\u7ec3\u9636\u6bb5\uff0c\u4e3a\u4e86\u91c7\u96c6\u8db3\u591f\u7684 rollout\uff0c\u901a\u5e38\u9700\u8981\u5e76\u53d1\u8fd0\u884c\u5927\u91cf Agent\uff0c\u56e0\u6b64 Agent\u6267\u884c\u672c\u8eab\u5c31\u4f1a\u6d88\u8017\u5927\u91cf CPU\u3001\u5185\u5b58\u548c\u8ba1\u7b97\u8d44\u6e90\u3002&nbsp;<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u73b0\u6709\u4e00\u4e9b Harnessed Agentic RL \u6846\u67b6\u901a\u5e38\u4f9d\u8d56 Modal Sandbox\u3001E2B \u7b49\u5546\u4e1a\u6c99\u76d2\uff08sandbox\uff09\u670d\u52a1\u6765\u627f\u8f7d\u8fd9\u4e9b Agent\u3002\u867d\u7136\u4f7f\u7528\u65b9\u4fbf\uff0c\u4f46\u5f53 rollout \u89c4\u6a21\u4e0d\u65ad\u6269\u5927\u65f6\uff0c\u5916\u90e8\u6c99\u76d2\u7684\u6301\u7eed\u8c03\u7528\u6210\u672c\u4e5f\u4f1a\u8fc5\u901f\u4e0a\u5347\u3002&nbsp;<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Agent Lightning v1.0&nbsp;\u5219\u53ef\u4ee5\u76f4\u63a5\u5c06&nbsp;Agent&nbsp;\u4f5c\u4e3a\u6807\u51c6&nbsp;Kubernetes&nbsp;\u4efb\u52a1\u8fd0\u884c\uff0c\u590d\u7528\u5df2\u6709\u7684\u81ea\u5efa\u96c6\u7fa4\u3001\u4e91\u4e0a Kubernetes \u6216\u672c\u5730\u57fa\u7840\u8bbe\u65bd\uff0c\u65e0\u9700\u4f9d\u8d56\u5916\u90e8\u5546\u4e1a\u6c99\u76d2\u3002&nbsp;<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u6837\u65e2\u80fd\u66f4\u9ad8\u6548\u5730\u5229\u7528\u5df2\u6709\u8ba1\u7b97\u8d44\u6e90\u3001\u964d\u4f4e\u5927\u89c4\u6a21 rollout \u7684\u989d\u5916\u6210\u672c\uff0c\u4e5f\u8ba9\u4ece Agent \u6267\u884c\u5230 RL \u8bad\u7ec3\u7684\u6574\u5957\u6d41\u7a0b\u4fdd\u6301\u5b8c\u5168\u5f00\u6e90\u3001\u53ef\u63a7\u4e14\u6613\u4e8e\u590d\u73b0\u3002&nbsp;<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"831\" height=\"233\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/10-4.jpg\" alt=\"diagram\" class=\"wp-image-1185176\" style=\"width:770px\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/10-4.jpg 831w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/10-4-300x84.jpg 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/10-4-768x215.jpg 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/10-4-240x67.jpg 240w\" sizes=\"auto, (max-width: 831px) 100vw, 831px\" \/><figcaption class=\"wp-element-caption\">\u56fe4\uff1aAgent&nbsp;Lightning v1.0\u7684Rollout&nbsp;Controller\u63d0\u4f9b\u4e86\u539f\u751f\u7684Kubernetes\u652f\u6301\uff0c\u5b83\u76f4\u63a5\u5c06Agent\u4f5c\u4e3a\u6807\u51c6Kubernetes\u4efb\u52a1\u8fd0\u884c\u3002&nbsp;<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n\n\n\n<h2 id=\"6-000\u6761\u8bad\u7ec3\u6570\u636e-\u6027\u80fd\u63d0\u534714-6\u4e2a\u767e\u5206\u70b9\" class=\"wp-block-heading\">6,000\u6761\u8bad\u7ec3\u6570\u636e\uff0c\u6027\u80fd\u63d0\u534714.6\u4e2a\u767e\u5206\u70b9&nbsp;<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u4e86\u9a8c\u8bc1 Harnessed Agentic RL \u7684\u5b9e\u9645\u6548\u679c\uff0c\u7814\u7a76\u5458\u4eec\u57fa\u4e8e SWE-smith + mini-SWE-agent + Qwen3.5-9B \u5efa\u7acb\u4e86\u4e00\u5957\u5b8c\u6574\u7684\u6570\u636e\u6e05\u6d17\u3001\u73af\u5883\u6784\u5efa\u3001\u5956\u52b1\u4f5c\u5f0a\uff08reward-hacking\uff09\u9632\u62a4\u548c RL&nbsp;\u8bad\u7ec3\u7ba1\u7ebf\u3002\u8bad\u7ec3\u96c6\u5305\u542b\u7ea6 6,000&nbsp;\u4e2a\u6837\u672c\uff0c\u65e0\u9700\u4f9d\u8d56\u5927\u89c4\u6a21\u8ba1\u7b97\u8d44\u6e90\u3002&nbsp;<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b9e\u9a8c\u7ed3\u679c\u663e\u793a\uff0c\u4ec5\u901a\u8fc7 RL&nbsp;\u8bad\u7ec3\uff0cQwen3.5-9B \u5728 SWE-bench Verified&nbsp;\u4e0a\u4ece&nbsp;41.8%&nbsp;\u63d0\u5347\u5230&nbsp;56.4%\uff0c\u7edd\u5bf9\u63d0\u534714.6\u4e2a\u767e\u5206\u70b9\u3002&nbsp;<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0e\u6b64\u540c\u65f6\uff0c\u4ee3\u7801&nbsp;Agent&nbsp;\u5b9e\u9a8c\u4e5f\u8fdb\u4e00\u6b65\u9a8c\u8bc1\u4e86\u524d\u6587\u5bf9\u201c\u4f18\u52bf\u503c\u8ba1\u7b97\u201d\u548c\u201c\u635f\u5931\u5f52\u4e00\u5316\u201d\u4e24\u5927\u6311\u6218\u7684\u5206\u6790\u3002\u76f8\u6bd4\u6837\u672c\u5c42\u7684\u5904\u7406\u65b9\u5f0f\uff0c\u91c7\u7528&nbsp;rollout&nbsp;\u5c42\u7ea7\u4f18\u52bf\u503c\uff08Rollout-level Advantage\uff09+&nbsp;rollout&nbsp;\u5c42\u7ea7\u5f52\u4e00\u5316\uff08Rollout-level Normalization\uff09\u80fd\u83b7\u5f97\u66f4\u9ad8\u7684\u9a8c\u8bc1\u5956\u52b1\uff0c\u540c\u65f6\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7684\u7b56\u7565\u71b5\uff08policy entropy\uff09\u4e5f\u66f4\u52a0\u7a33\u5b9a\u3002&nbsp;<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"831\" height=\"303\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/10-5.png\" alt=\"chart, line chart\" class=\"wp-image-1185177\" style=\"width:770px\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/10-5.png 831w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/10-5-300x109.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/10-5-768x280.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/10-5-240x88.png 240w\" sizes=\"auto, (max-width: 831px) 100vw, 831px\" \/><figcaption class=\"wp-element-caption\">\u56fe5\uff1aQwen3.5-9B\u5728SWE-smith\u9a8c\u8bc1\u96c6\u4e0a\u7684\u901a\u8fc7\u7387\u548c\u7b56\u7565\u71b5\u3002&nbsp;<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n\n\n\n<h2 id=\"\u8ba9-agent-rl-\u66f4\u8d34\u8fd1\u771f\u5b9e\u4e16\u754c\" class=\"wp-block-heading\">\u8ba9 Agent RL \u66f4\u8d34\u8fd1\u771f\u5b9e\u4e16\u754c&nbsp;<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Agent Lightning v1.0&nbsp;\u805a\u7126\u4e8e\u4e00\u4e2a\u7b80\u5355\u5374\u91cd\u8981\u7684\u95ee\u9898\uff1a\u5982\u679c\u672a\u6765\u7684&nbsp;Agent&nbsp;\u8fd0\u884c\u5728\u771f\u5b9e\u7684&nbsp;Agent Harness&nbsp;\u4e4b\u4e2d\uff0c\u90a3\u4e48\u4e3a\u4ec0\u4e48\u8bad\u7ec3\u65f6\u4e0d\u80fd\u76f4\u63a5\u4f7f\u7528\u540c\u6837\u7684&nbsp;Harness\uff1f&nbsp;<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u56f4\u7ed5\u8fd9\u4e00\u7406\u5ff5\uff0cAgent Lightning v1.0 \u8ba9\u771f\u5b9e\u90e8\u7f72\u73af\u5883\u4e2d\u7684 Agent Harness \u80fd\u591f\u76f4\u63a5\u53c2\u4e0e\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\uff0c\u4ece\u800c\u4fdd\u7559\u5de5\u5177\u8c03\u7528\u3001\u4e0a\u4e0b\u6587\u7ba1\u7406\u3001\u63a7\u5236\u6d41\u4ee5\u53ca\u6267\u884c\u73af\u5883\u3002\u540c\u65f6\uff0c\u8fd9\u4e00\u8bad\u7ec3\u8303\u5f0f\u4e5f\u5e26\u6765\u4e86\u65b0\u7684\u7814\u7a76\u95ee\u9898\uff0c\u5305\u62ec\u91cd\u65b0\u5206\u8bcd\u3001\u52a8\u6001\u6837\u672c\u6570\u91cf\u3001\u4f18\u52bf\u503c\u8ba1\u7b97\u3001\u635f\u5931\u5f52\u4e00\u5316\u548c\u8d44\u6e90\u8c03\u5ea6\u7b49\u6311\u6218\u3002Agent Lightning v1.0 \u5bf9\u8fd9\u4e9b\u95ee\u9898\u8fdb\u884c\u4e86\u7cfb\u7edf\u5316\u68b3\u7406\uff0c\u5e76\u4ee5\u7ea6 3,500 \u884c\u4ee3\u7801\u5b9e\u73b0\u4e86\u4e00\u5957\u8f7b\u91cf\u3001\u900f\u660e\u3001\u53ef\u590d\u73b0\u7684\u89e3\u51b3\u65b9\u6848\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7814\u7a76\u5458\u4eec\u5e0c\u671b\uff0c\u968f\u7740 Agent \u7cfb\u7edf\u4e0d\u65ad\u8d70\u5411\u771f\u5b9e\u5e94\u7528\u573a\u666f\uff0c\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u4e5f\u80fd\u591f\u9010\u6e10\u6446\u8131\u7406\u60f3\u5316\u73af\u5883\u7684\u9650\u5236\u3002Harnessed Agentic RL \u7684\u63a2\u7d22\uff0c\u6b63\u662f\u5728\u8fd9\u4e00\u65b9\u5411\u4e0a\u7684\u4e00\u6b21\u5c1d\u8bd5\uff0c\u5e76\u6709\u671b\u8fdb\u4e00\u6b65\u7f29\u5c0f Agent \u7814\u7a76\u4e0e\u73b0\u5b9e\u5e94\u7528\u4e4b\u95f4\u7684\u8ddd\u79bb\u3002&nbsp;<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Agent Lightning v1.0: Towards Harnessed Agentic RL<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bba\u6587\u94fe\u63a5\uff1a<a class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" rel=\"noopener noreferrer\" target=\"_blank\" href=\"https:\/\/arxiv.org\/pdf\/2608.17528\">https:\/\/arxiv.org\/pdf\/2608.17528<span class=\"sr-only\"> (opens in new tab)<\/span><\/a><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9879\u76ee\u9875\u9762\uff1a<a class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" rel=\"noopener noreferrer\" target=\"_blank\" href=\"https:\/\/github.com\/microsoft\/agent-lightning&nbsp;&nbsp;\">https:\/\/github.com\/microsoft\/agent-lightning&nbsp;&nbsp;<span class=\"sr-only\"> (opens in new tab)<\/span><\/a><\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u968f\u7740 AI Agent \u4ece\u5355\u7eaf\u7684\u6a21\u578b\u9010\u6e10\u6f14\u5316\u4e3a\u7531\u6a21\u578b\u3001\u5de5\u5177\u548c\u6267\u884c\u73af\u5883\u5171\u540c\u7ec4\u6210\u7684\u590d\u6742\u7cfb\u7edf\uff0cAgent \u7684\u80fd\u529b\u8d8a\u6765\u8d8a\u4f9d\u8d56\u4e8e\u6a21\u578b\u4e4b\u5916\u7684&nbsp;Agent&nbsp;Harness\uff08Agent\u8fd0\u884c\u6846\u67b6\uff09\u3002 \u7136\u800c\uff0c\u5f53\u6211\u4eec\u5e0c\u671b\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u8fdb\u4e00\u6b65\u63d0\u5347 Agent \u80fd\u529b\u65f6\uff0c\u4e00\u4e2a\u73b0\u5b9e\u95ee\u9898\u51fa\u73b0\u4e86\uff1a\u8bb8\u591a\u73b0\u6709 Agent RL \u7cfb\u7edf\u8981\u6c42\u5f00\u53d1\u8005\u5c06 Agent \u91cd\u65b0\u5b9e\u73b0\u5230\u8bad\u7ec3\u6846\u67b6\u5185\u90e8\u3002\u8fd9\u4e0d\u4ec5\u589e\u52a0\u4e86\u5de5\u7a0b\u6210\u672c\uff0c\u4e5f\u8ba9\u8bad\u7ec3\u73af\u5883\u4e0e\u771f\u5b9e\u90e8\u7f72\u73af\u5883\u9010\u6e10\u8131\u8282\u3002 \u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86&nbsp;Harnessed Agentic RL&nbsp;\u8bad\u7ec3\u8303\u5f0f\uff0c\u5e76\u5f00\u6e90\u4e86\u5168\u9762\u91cd\u6784\u7684 Agent Lightning v1.0&nbsp;\u3002\u76f8\u6bd4\u539f\u7248\u672c\u7684Agent&nbsp;Lightning\uff0cv1.0&nbsp;\u66f4\u5f3a\u8c03\u8f7b\u91cf\u3001\u771f\u5b9e\u7684&nbsp;Harness&nbsp;\u96c6\u6210\u4ee5\u53ca\u5b8c\u6574\u53ef\u590d\u73b0\u7684&nbsp;Agent RL&nbsp;\u8bad\u7ec3\u6d41\u7a0b\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c\u5728&nbsp;Agent Lightning v1.0&nbsp;\u4e2d\uff0c\u90e8\u7f72\u65f6\u4f7f\u7528\u4ec0\u4e48 Agent Harness\uff0c\u8bad\u7ec3\u65f6\u5c31\u76f4\u63a5\u8ba9\u8fd9\u4e2a Harness \u53c2\u4e0e\u5f3a\u5316\u5b66\u4e60\u3002 \u4e3a\u4e86\u8ba9 Agent RL \u66f4\u8d34\u8fd1\u771f\u5b9e\u4e16\u754c\u4e2d\u7684 Agent \u7cfb\u7edf\uff0cAgent Lightning v1.0 \u56f4\u7ed5&nbsp;Harnessed Agentic RL \u8fdb\u884c\u4e86\u5168\u9762\u91cd\u6784\uff0c\u5728\u7cfb\u7edf\u8bbe\u8ba1\u3001\u5de5\u7a0b\u5b9e\u8df5\u548c\u8bad\u7ec3\u6548\u679c\u7b49\u65b9\u9762\u90fd\u5e26\u6765\u4e86\u591a\u9879\u5173\u952e\u6539\u8fdb\u3002&nbsp; \u8bba\u6587\u4fe1\u606f\u5df2\u6574\u7406\u81f3\u6587\u672b\uff0c\u6b22\u8fce\u70b9\u51fb\u76f8\u5173\u94fe\u63a5\uff0c\u4e86\u89e3\u66f4\u591a\u6280\u672f\u8be6\u60c5\u3002 \u4f20\u7edf Agentic RL \u901a\u5e38\u5047\u8bbe\u8bad\u7ec3\u6846\u67b6\u81ea\u5df1\u62e5\u6709\u5b8c\u6574\u7684\u73af\u5883\u4ea4\u4e92\u5faa\u73af\u3002\u4f8b\u5982\uff0c\u5728\u5178\u578b\u7684&nbsp;ReAct&nbsp;Agent&nbsp;\u4e2d\uff0c\u6a21\u578b\u751f\u6210\u52a8\u4f5c\uff08action\uff09\uff0c\u73af\u5883\u8fd4\u56de\u89c2\u5bdf\u7ed3\u679c\uff08observation\uff09\uff0c\u89c2\u5bdf\u7ed3\u679c\u88ab\u8ffd\u52a0\u5230\u5df2\u6709\u4e0a\u4e0b\u6587\u4e2d\uff0c\u968f\u540e\u6a21\u578b\u7ee7\u7eed\u751f\u6210\u4e0b\u4e00\u6b65\u52a8\u4f5c\u3002\u6574\u4e2a rollout&nbsp;\u56e0\u6b64\u5929\u7136\u5bf9\u5e94\u4e00\u6761\u8fde\u7eed\u7684 token&nbsp;\u8f68\u8ff9\u3002\u65e9\u671f\u7684&nbsp;verl\u3001AReaL\u3001slime&nbsp;\u7b49\u5f3a\u5316\u5b66\u4e60\u7cfb\u7edf\u57fa\u672c\u6cbf\u7528\u4e86\u8fd9\u4e00\u8bbe\u8ba1\u3002\u5982\u679c\u60f3\u8bad\u7ec3\u4e00\u4e2a Agent\uff0c\u5f00\u53d1\u8005\u901a\u5e38\u9700\u8981\u628a Agent\u5faa\u73af\u91cd\u65b0\u5b9e\u73b0\u5230 RL \u6846\u67b6\u5185\u90e8\u3002 \u4f46\u73b0\u5b9e\u4e2d\u7684 Agent Harness \u5df2\u7ecf\u8d8a\u6765\u8d8a\u590d\u6742\u3002\u65e0\u8bba\u662fmini-SWE-agent\u3001OpenHands\u3001OpenCode\u3001ClaudeCode\u3001Codex&nbsp;\u7b49\u4ee3\u7801&nbsp;Agent \uff0c\u8fd8\u662f\u5404\u79cd\u901a\u7528 &nbsp;Agent&nbsp;\u7cfb\u7edf\uff0c\u90fd\u62e5\u6709\u72ec\u7acb\u7684\u4e0a\u4e0b\u6587\u7ba1\u7406\u3001\u5de5\u5177\u534f\u8bae\u3001\u6267\u884c\u903b\u8f91\u548c\u8f6f\u4ef6\u4f9d\u8d56\u3002\u4e3a\u4e86 [&hellip;]<\/p>\n","protected":false},"author":44093,"featured_media":1185178,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"_classifai_error":"","msr-content-parent":1012650,"msr_hide_image_in_river":0,"footnotes":""},"research-area":[13561,13556,13560,13547],"msr-locale":[268881],"msr-post-option":[],"class_list":["post-1185172","msr-blog-post","type-msr-blog-post","status-publish","has-post-thumbnail","hentry","msr-research-area-algorithms","msr-research-area-artificial-intelligence","msr-research-area-programming-languages-software-engineering","msr-research-area-systems-and-networking","msr-locale-zh_cn"],"msr_assoc_parent":{"id":1012650,"type":"lab"},"_links":{"self":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1185172","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post"}],"about":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-blog-post"}],"author":[{"embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/users\/44093"}],"version-history":[{"count":3,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1185172\/revisions"}],"predecessor-version":[{"id":1185417,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1185172\/revisions\/1185417"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media\/1185178"}],"wp:attachment":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1185172"}],"wp:term":[{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1185172"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1185172"},{"taxonomy":"msr-post-option","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-post-option?post=1185172"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}