[{"data":1,"prerenderedAt":4},["ShallowReactive",2],{"readme:jit":3},"\u003Cdiv align=\"center\">\u003Cimg src=\"https:\u002F\u002Fraw.githubusercontent.com\u002Fbingreeky\u002Fjit\u002FHEAD\u002Fassets\u002Fjit-agent-logo.png\" alt=\"JIT-Agent\" width=\"380\" \u002F>\u003Cp>\u003Cstrong>Scaling Harness Intelligence via \u003Cem>Just-in-Time\u003C\u002Fem> Harness Evolution\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cp>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fbingreeky\u002FJIT\" rel=\"nofollow ugc noopener\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FGitHub-bingreeky%2FJIT-181717?style=flat-square&logo=github\" alt=\"GitHub\" \u002F>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.25593\" rel=\"nofollow ugc noopener\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2608.25593-b31b1b.svg?style=flat-square\" alt=\"arXiv\" \u002F>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002FJIT-Agent\" rel=\"nofollow ugc noopener\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F%F0%9F%A4%97%20Hugging%20Face-JIT--Agent-ffbd45?style=flat-square\" alt=\"Hugging Face\" \u002F>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fbingreeky\u002Fjit\u002Fblob\u002FHEAD\u002FLICENSE\" rel=\"nofollow ugc noopener\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLicense-see%20LICENSE-blue?style=flat-square\" alt=\"License\" \u002F>\u003C\u002Fa>\u003C\u002Fp>\n\u003C\u002Fdiv>\u003Ch2>What is JIT-Agent?\u003C\u002Fh2>\n\u003Cp>\u003Cstrong>JIT-Agent is a compact meta-agent that writes your agent harness on the fly.\u003C\u002Fstrong> Instead of\nprecompiling one general-purpose scaffold and hoping it transfers, JIT-Agent takes a task\nspec, a protocol, a tool\u002Fskill registry, and a few retrieved prior harnesses, and emits an\n\u003Cstrong>executable, task-specific harness that wraps any off-the-shelf agentic LLM\u003C\u002Fstrong> —\n\u003Cem>\u003Cstrong>Model-as-a-Harness\u003C\u002Fstrong>\u003C\u002Fem>.\u003C\u002Fp>\n\u003Cdiv align=\"center\">\n\u003Cimg src=\"https:\u002F\u002Fraw.githubusercontent.com\u002Fbingreeky\u002Fjit\u002FHEAD\u002Fassets\u002Fmethod_overview.png\" alt=\"Overview of JIT-Agent\" width=\"100%\" \u002F>\n\u003C\u002Fdiv>\u003Cdiv align=\"center\">\u003Csub>\u003Cb>Overview of JIT-Agent.\u003C\u002Fb> Given a task, JIT-Agent composes a\nproblem-specific agent harness by selecting and instantiating four modules: memory,\nplanning, action, and capability. Different task structures therefore induce distinct\nexecutable protocols and state organizations.\u003C\u002Fsub>\u003C\u002Fdiv>\u003Cp>Every harness is factored into \u003Cstrong>four modules — memory, planning, action, capability\norchestration\u003C\u002Fstrong> — implemented against the shared interfaces in\n\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fbingreeky\u002Fjit\u002Fblob\u002FHEAD\u002Fharness_factory\u002F\" rel=\"nofollow ugc noopener\">\u003Cstrong>HarnessFactory\u003C\u002Fstrong>\u003C\u002Fa>, so generation means \u003Cstrong>emitting structured code\nrather than free-form agent programs\u003C\u002Fstrong>. As traces and feedback come back, JIT-Agent revises\nthe harness and updates the archive: \u003Cstrong>harnesses keep improving at test time while the\ngenerator itself stays frozen.\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cdiv align=\"center\">\n\u003Cimg src=\"https:\u002F\u002Fraw.githubusercontent.com\u002Fbingreeky\u002Fjit\u002FHEAD\u002Fassets\u002Fleaderboard.png\" alt=\"JIT-Agent leaderboard across four representative agent benchmarks\" width=\"100%\" \u002F>\n\u003C\u002Fdiv>\u003Cp>\u003Cstrong>Results.\u003C\u002Fstrong> The resulting \u003Cstrong>JIT-Agent-27B\u003C\u002Fstrong> lifts a wide range of backbone agents across\ndeep research, daily work, planning, and workspace tasks.\u003C\u002Fp>\n\u003Cblockquote>\n\u003Cp>\u003Cstrong>Building the scaffold turns out to be a trainable, transferable axis of agent\nintelligence — orthogonal to scaling the base model.\u003C\u002Fstrong>\u003C\u002Fp>\n\u003C\u002Fblockquote>\n\u003Chr \u002F>\n\u003Ch2>Repository layout\u003C\u002Fh2>\n\u003Ctable>\n\u003Cthead>\n\u003Ctr>\n\u003Cth>Directory\u003C\u002Fth>\n\u003Cth>What it holds\u003C\u002Fth>\n\u003C\u002Ftr>\n\u003C\u002Fthead>\n\u003Ctbody>\u003Ctr>\n\u003Ctd>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fbingreeky\u002Fjit\u002Fblob\u002FHEAD\u002Fjit\u002F\" rel=\"nofollow ugc noopener\">\u003Ccode>jit\u002F\u003C\u002Fcode>\u003C\u002Fa>\u003C\u002Ftd>\n\u003Ctd>The meta agent: generation \u002F repair prompts, best-of-N selection\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fbingreeky\u002Fjit\u002Fblob\u002FHEAD\u002Fscripts\u002F\" rel=\"nofollow ugc noopener\">\u003Ccode>scripts\u002F\u003C\u002Fcode>\u003C\u002Fa>\u003C\u002Ftd>\n\u003Ctd>The agent kernel, tools, models, evaluation engine, and the two runners\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fbingreeky\u002Fjit\u002Fblob\u002FHEAD\u002Fharness_factory\u002F\" rel=\"nofollow ugc noopener\">\u003Ccode>harness_factory\u002F\u003C\u002Fcode>\u003C\u002Fa>\u003C\u002Ftd>\n\u003Ctd>Hand-written harness implementations and their design write-ups\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fbingreeky\u002Fjit\u002Fblob\u002FHEAD\u002Fbenchmark\u002F\" rel=\"nofollow ugc noopener\">\u003Ccode>benchmark\u002F\u003C\u002Fcode>\u003C\u002Fa>\u003C\u002Ftd>\n\u003Ctd>One adapter, config and evaluator per benchmark\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fbingreeky\u002Fjit\u002Fblob\u002FHEAD\u002Fdataset\u002F\" rel=\"nofollow ugc noopener\">\u003Ccode>dataset\u002F\u003C\u002Fcode>\u003C\u002Fa>\u003C\u002Ftd>\n\u003Ctd>The benchmark data itself\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftbody>\u003C\u002Ftable>\n\u003Cp>Each directory has its own README with the details.\u003C\u002Fp>\n\u003Ch2>Setup\u003C\u002Fh2>\n\u003Cp>\u003Cstrong>1. Clone the repository\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cpre>\u003Ccode class=\"language-bash\">git clone https:\u002F\u002Fgithub.com\u002Fbingreeky\u002FJIT.git\ncd JIT\n\u003C\u002Fcode>\u003C\u002Fpre>\n\u003Cp>\u003Cstrong>2. Environment\u003C\u002Fstrong> (Python 3.11)\u003C\u002Fp>\n\u003Cpre>\u003Ccode class=\"language-bash\">conda env create -f environment.yml && conda activate jit\n\u003C\u002Fcode>\u003C\u002Fpre>\n\u003Cp>or, in an existing environment: \u003Ccode>pip install -r requirements.txt\u003C\u002Fcode>. Serving a local meta\nmodel (vLLM\u002FSGLang + torch) is deliberately not included — the pipeline only ever talks\nHTTP to it.\u003C\u002Fp>\n\u003Cp>\u003Cstrong>3. Credentials\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cpre>\u003Ccode class=\"language-bash\">cp .env.example .env   # then fill it in\n\u003C\u002Fcode>\u003C\u002Fpre>\n\u003Cp>Anything already exported in the shell wins over \u003Ccode>.env\u003C\u002Fcode>, and every model role can also be\noverridden per run on the command line.\u003C\u002Fp>\n\u003Ctable>\n\u003Cthead>\n\u003Ctr>\n\u003Cth>Group\u003C\u002Fth>\n\u003Cth>Keys\u003C\u002Fth>\n\u003Cth>Used for\u003C\u002Fth>\n\u003C\u002Ftr>\n\u003C\u002Fthead>\n\u003Ctbody>\u003Ctr>\n\u003Ctd>Execution model\u003C\u002Ftd>\n\u003Ctd>\u003Ccode>OPENAI_API_BASE\u003C\u002Fcode>, \u003Ccode>OPENAI_API_KEY\u003C\u002Fcode>, \u003Ccode>EXEC_MODEL\u003C\u002Fcode>\u003C\u002Ftd>\n\u003Ctd>runs the generated harness's agent loop\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Judge model\u003C\u002Ftd>\n\u003Ctd>\u003Ccode>JUDGE_MODEL\u003C\u002Fcode>, optional \u003Ccode>JUDGE_API_*\u003C\u002Fcode>\u003C\u002Ftd>\n\u003Ctd>grades produced artifacts (falls back to the execution endpoint)\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Meta model\u003C\u002Ftd>\n\u003Ctd>\u003Ccode>META_MODEL\u003C\u002Fcode>, \u003Ccode>META_API_BASE\u003C\u002Fcode>, \u003Ccode>META_API_KEY\u003C\u002Fcode>, \u003Ccode>META_TOKENIZER\u003C\u002Fcode>\u003C\u002Ftd>\n\u003Ctd>writes the harness (JIT pipeline only)\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Tools\u003C\u002Ftd>\n\u003Ctd>\u003Ccode>SERPER_API_KEY\u003C\u002Fcode>, \u003Ccode>JINA_API_KEY\u003C\u002Fcode>\u003C\u002Ftd>\n\u003Ctd>`web_s\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftbody>\u003C\u002Ftable>\n",1788652555354]