[{"data":1,"prerenderedAt":4},["ShallowReactive",2],{"readme:reef":3},"\u003Cdiv align=\"center\">\u003Cpicture>\n  \u003Csource media=\"(prefers-color-scheme: dark)\" srcset=\"docs\u002Fassets\u002Freef-logo-dark.svg\">\u003C\u002Fsource>\n  \u003Cimg src=\"https:\u002F\u002Fraw.githubusercontent.com\u002FHuman-Agent-Society\u002Freef\u002FHEAD\u002Fdocs\u002Fassets\u002Freef-logo-light.svg\" alt=\"Reef\" width=\"220\" \u002F>\n\u003C\u002Fpicture>\u003Ch3>Continual learning infra for self-improving agents\u003C\u002Fh3>\u003Cp>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FHuman-Agent-Society\u002Freef\u002Factions\u002Fworkflows\u002Fci.yml\" rel=\"nofollow ugc noopener\">\u003Cimg src=\"https:\u002F\u002Fgithub.com\u002FHuman-Agent-Society\u002Freef\u002Factions\u002Fworkflows\u002Fci.yml\u002Fbadge.svg\" alt=\"CI\" \u002F>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fpypi.org\u002Fproject\u002Freef-infra\u002F\" rel=\"nofollow ugc noopener\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fpypi\u002Fv\u002Freef-infra?label=PyPI%3A%20reef-infra&logo=pypi&logoColor=white\" alt=\"PyPI package: reef-infra\" \u002F>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FHuman-Agent-Society\u002Freef\u002Fblob\u002FHEAD\u002Fpyproject.toml\" rel=\"nofollow ugc noopener\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fpython-3.10%2B-3776AB?logo=python&logoColor=white\" alt=\"Python\" \u002F>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FHuman-Agent-Society\u002Freef\u002Fblob\u002FHEAD\u002FLICENSE\" rel=\"nofollow ugc noopener\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Flicense-Apache--2.0-green\" alt=\"License\" \u002F>\u003C\u002Fa>\u003C\u002Fp>\n\u003Cp>English | \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FHuman-Agent-Society\u002Freef\u002Fblob\u002FHEAD\u002FREADME.zh.md\" rel=\"nofollow ugc noopener\">中文\u003C\u002Fa>\u003C\u002Fp>\n\u003Cdiv align=\"left\">\u003Cp>Reef is the first open-source infrastructure for continual self-improving agents.\nIt connects agent inference, feedback, learning, and versioned delivery. Use it\nto train model weights with Slime and SGLang, or improve an agent's harness, including its prompts, rules, and skills.\u003C\u002Fp>\n\u003C\u002Fdiv>\u003Cp>\u003Cstrong>\u003Ca href=\"https:\u002F\u002Freefinfra.ai\u002Fdocs\u002Fgetting-started\u002Fquickstart\u002F\" rel=\"nofollow ugc noopener\">Get started\u003C\u002Fa> |\n\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FHuman-Agent-Society\u002Freef\u002Fissues\u002F25\" rel=\"nofollow ugc noopener\">Roadmap\u003C\u002Fa> |\n\u003Ca href=\"https:\u002F\u002Fx.com\u002Fao_qu18465\u002Fstatus\u002F2094867930081337730\" rel=\"nofollow ugc noopener\">Launch post\u003C\u002Fa> |\n\u003Ca href=\"https:\u002F\u002Fdiscord.gg\u002F5y8e5f937k\" rel=\"nofollow ugc noopener\">Join Discord\u003C\u002Fa> |\n\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FHuman-Agent-Society\u002Freef\u002Fblob\u002FHEAD\u002Fdocs\u002Fcommunity\u002Fwechat.md\" rel=\"nofollow ugc noopener\">Join WeChat Group\u003C\u002Fa>\u003C\u002Fstrong>\u003C\u002Fp>\n\u003C\u002Fdiv>\u003Ch2>When to use Reef\u003C\u002Fh2>\n\u003Cp>Use Reef when you want your agent to keep improving simply by learning from how you interact with your agent.\u003C\u002Fp>\n\u003Ctable>\n\u003Cthead>\n\u003Ctr>\n\u003Cth>Your goal\u003C\u002Fth>\n\u003Cth>Learning path\u003C\u002Fth>\n\u003Cth>What you need\u003C\u002Fth>\n\u003C\u002Ftr>\n\u003C\u002Fthead>\n\u003Ctbody>\u003Ctr>\n\u003Ctd>Keep getting stronger model designed for you\u003C\u002Ftd>\n\u003Ctd>Model weight training\u003C\u002Ftd>\n\u003Ctd>A trainable model, a supported GPU stack, and feedback your recipe can use\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Get your harness to self-improve\u003C\u002Ftd>\n\u003Ctd>Harness optimization\u003C\u002Ftd>\n\u003Ctd>A model endpoint, representative tasks, and an evaluator; no local training GPUs\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Scientific discoveries\u003C\u002Ftd>\n\u003Ctd>Test-time training\u003C\u002Ftd>\n\u003Ctd>An execution environment, a correctness checker, and a measurable objective\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftbody>\u003C\u002Ftable>\n\u003Ch2>How Reef fits your stack\u003C\u002Fh2>\n\u003Ctable>\n\u003Cthead>\n\u003Ctr>\n\u003Cth>Ability\u003C\u002Fth>\n\u003Cth align=\"center\">Inference engine (vLLM, SGLang, …)\u003C\u002Fth>\n\u003Cth align=\"center\">RL training framework (Slime, veRL, AReaL, …)\u003C\u002Fth>\n\u003Cth align=\"center\">\u003Cstrong>Reef\u003C\u002Fstrong>\u003C\u002Fth>\n\u003C\u002Ftr>\n\u003C\u002Fthead>\n\u003Ctbody>\u003Ctr>\n\u003Ctd>Serves live traffic\u003C\u002Ftd>\n\u003Ctd align=\"center\">✅\u003C\u002Ftd>\n\u003Ctd align=\"center\">❌\u003C\u002Ftd>\n\u003Ctd align=\"center\">✅\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Trains weights\u003C\u002Ftd>\n\u003Ctd align=\"center\">❌\u003C\u002Ftd>\n\u003Ctd align=\"center\">✅\u003C\u002Ftd>\n\u003Ctd align=\"center\">✅\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Version management\u003C\u002Ftd>\n\u003Ctd align=\"center\">❌\u003C\u002Ftd>\n\u003Ctd align=\"center\">❌\u003C\u002Ftd>\n\u003Ctd align=\"center\">✅\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Stays live through updates\u003C\u002Ftd>\n\u003Ctd align=\"center\">❌\u003C\u002Ftd>\n\u003Ctd align=\"center\">❌\u003C\u002Ftd>\n\u003Ctd align=\"center\">✅\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Evolves beyond weights (skills, harness)\u003C\u002Ftd>\n\u003Ctd align=\"center\">❌\u003C\u002Ftd>\n\u003Ctd align=\"center\">❌\u003C\u002Ftd>\n\u003Ctd align=\"center\">✅\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftbody>\u003C\u002Ftable>\n\u003Ch2>How it works\u003C\u002Fh2>\n\u003Cdiv align=\"center\">\n\u003Cpicture>\n  \u003Csource media=\"(prefers-color-scheme: dark)\" srcset=\"docs\u002Fassets\u002Floop-animation-dark.svg\">\u003C\u002Fsource>\n  \u003Cimg src=\"https:\u002F\u002Fraw.githubusercontent.com\u002FHuman-Agent-Society\u002Freef\u002FHEAD\u002Fdocs\u002Fassets\u002Floop-animation-light.svg\" alt=\"Reef serves requests, records feedback, produces updates, and commits accepted updates to a version history.\" width=\"76%\" \u002F>\n\u003C\u002Fpicture>\n\u003C\u002Fdiv>\u003Cp>Reef processes each learning cycle in four steps. The table also shows which\nmodules implement each step.\u003C\u002Fp>\n\u003Ctable>\n\u003Cthead>\n\u003Ctr>\n\u003Cth>Step\u003C\u002Fth>\n\u003Cth>What happens\u003C\u002Fth>\n\u003Cth>Where it lives\u003C\u002Fth>\n\u003C\u002Ftr>\n\u003C\u002Fthead>\n\u003Ctbody>\u003Ctr>\n\u003Ctd>\u003Cstrong>1 · Serve\u003C\u002Fstrong>\u003C\u002Ftd>\n\u003Ctd>Serve agent requests and record interactions.\u003C\u002Ftd>\n\u003Ctd>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FHuman-Agent-Society\u002Freef\u002Fblob\u002FHEAD\u002Freef\u002Fservice\" rel=\"nofollow ugc noopener\">\u003Ccode>service\u002F\u003C\u002Fcode>\u003C\u002Fa> — agent requests and interaction records\u003Cbr \u002F>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FHuman-Agent-Society\u002Freef\u002Fblob\u002FHEAD\u002Freef\u002Fruntime\" rel=\"nofollow ugc noopener\">\u003Ccode>runtime\u002F\u003C\u002Fcode>\u003C\u002Fa> — inference and artifact updates\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Cstrong>2 · Observe\u003C\u002Fstrong>\u003C\u002Ftd>\n\u003Ctd>Match feedback to recorded interactions.\u003C\u002Ftd>\n\u003Ctd>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FHuman-Agent-Society\u002Freef\u002Fblob\u002FHEAD\u002Freef\u002Frecords.py\" rel=\"nofollow ugc noopener\">\u003Ccode>records.py\u003C\u002Fcode>\u003C\u002Fa> — stored interactions and feedback\u003Cbr \u002F>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FHuman-Agent-Society\u002Freef\u002Fblob\u002FHEAD\u002Freef\u002Ftrain\u002Fprocessors\" rel=\"nofollow ugc noopener\">\u003Ccode>train\u002Fprocessors\u002F\u003C\u002Fcode>\u003C\u002Fa> — feedback matching and eligibility\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Cstrong>3 · Grow\u003C\u002Fstrong>\u003C\u002Ftd>\n\u003Ctd>Produce an update from eligible records.\u003C\u002Ftd>\n\u003Ctd>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FHuman-Agent-Society\u002Freef\u002Fblob\u002FHEAD\u002Freef\u002Frecipe\" rel=\"nofollow ugc noopener\">\u003Ccode>recipe\u002F\u003C\u002Fcode>\u003C\u002Fa> — recipe integration\u003Cbr \u002F>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FHuman-Agent-Society\u002Freef\u002Fblob\u002FHEAD\u002Freef\u002Ftrain\" rel=\"nofollow ugc noopener\">\u003Ccode>train\u002F\u003C\u002Fcode>\u003C\u002Fa> — batches and update jobs\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Cstrong>4 · Commit\u003C\u002Fstrong>\u003C\u002Ftd>\n\u003Ctd>Apply the configured selection policy and publish accepted updates.\u003C\u002Ftd>\n\u003Ctd>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FHuman-Agent-Society\u002Freef\u002Fblob\u002FHEAD\u002Freef\u002Ftrain\u002Fevaluation\" rel=\"nofollow ugc noopener\">\u003Ccode>train\u002Fevaluation\u002F\u003C\u002Fcode>\u003C\u002Fa> — candidate evaluation\u003Cbr \u002F>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FHuman-Agent-Society\u002Freef\u002Fblob\u002FHEAD\u002Freef\u002Fartifact\" rel=\"nofollow ugc noopener\">\u003Ccode>artifact\u002F\u003C\u002Fcode>\u003C\u002Fa> — version history\u003Cbr \u002F>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FHuman-Agent-Society\u002Freef\u002Fblob\u002FHEAD\u002Freef\u002Fsurface\" rel=\"nofollow ugc noopener\">\u003Ccode>surface\u002F\u003C\u002Fcode>\u003C\u002Fa> — artifact delivery\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftbody>\u003C\u002Ftable>\n\u003Ch2>Installation\u003C\u002Fh2>\n\u003Cblockquote>\n\u003Cp>💡 \u003Cstrong>Note\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cp>Reef's artifact and checkpoint functionality requires the \u003Ccode>git-lfs\u003C\u002Fcode> system\npackage. Reef initializes Git LFS locally for its artifact repositories.\u003C\u002Fp>\n\u003C\u002Fblockquote>\n\u003Cp>We recommend [uv](\u003Ca href=\"https:\u002F\u002Fdocs.ast\" rel=\"nofollow ugc noopener\">https:\u002F\u002Fdocs.ast\u003C\u002Fa>\u003C\u002Fp>\n",1789121864988]