[{"data":1,"prerenderedAt":4},["ShallowReactive",2],{"readme:harness2":3},"\u003Ch3>\n  \u003Cimg src=\"https:\u002F\u002Fraw.githubusercontent.com\u002Fgoogle-research\u002Fharness2\u002FHEAD\u002Fassets\u002FHarness2-wordmark.png\" alt=\"Harness²\" width=\"360\" \u002F>\u003Cbr \u002F>\n  Recursive Agent Harnessing for an Open World\n\u003C\u002Fh3>\u003Cp align=\"center\">\n  \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fpdf\u002F2610.14224\" rel=\"nofollow ugc noopener\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FPaper-arXiv-EA4335?logo=arxiv&logoColor=white\" alt=\"Paper\" \u002F>\u003C\u002Fa>\n  \u003Ca href=\"https:\u002F\u002Fharness2.github.io\" rel=\"nofollow ugc noopener\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FWebsite-harness2.github.io-4285F4?logo=googlechrome&logoColor=white\" alt=\"Website\" \u002F>\u003C\u002Fa>\n  \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fgoogle-research\u002Fharness2\u002Fblob\u002FHEAD\u002Fdocs\u002FBENCHMARKS.md\" rel=\"nofollow ugc noopener\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FDocs-Benchmark%20guides-FBBC04\" alt=\"Benchmark guides\" \u002F>\u003C\u002Fa>\n  \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fgoogle-research\u002Fharness2\u002Fblob\u002FHEAD\u002FLICENSE\" rel=\"nofollow ugc noopener\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLicense-Apache%202.0-34A853\" alt=\"License\" \u002F>\u003C\u002Fa>\n\u003C\u002Fp>\u003Cp align=\"center\">\n  \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fgoogle-research\u002Fharness2\u002Fblob\u002FHEAD\u002Fdocs\u002FBENCHMARKS.md#jobbench\" rel=\"nofollow ugc noopener\">\n    \u003Cimg src=\"https:\u002F\u002Fraw.githubusercontent.com\u002Fgoogle-research\u002Fharness2\u002FHEAD\u002Fassets\u002Fharness2-setup.gif\" alt=\"Install, set up, and run JobBench: uv pip install -e '.[render]', harness2-setup jb, then harness2 --bench jb --k 1 --mode parallel. Click for prerequisites and commands.\" width=\"640\" \u002F>\n  \u003C\u002Fa>\n\u003C\u002Fp>\u003Cp>A \u003Cstrong>harness\u003C\u002Fstrong> is the system around a model that enables it to act as an agent:\nprompts, tools, memory, skills, orchestration, and more. \u003Cstrong>Harnessing\u003C\u002Fstrong> is the\nprocess of optimizing that system.\u003C\u002Fp>\n\u003Cp>Harness² is a plug-and-play framework for \u003Cstrong>recursive agent harnessing\u003C\u002Fstrong> in an\nopen world, where agents encounter tasks unseen during harness design. It\nconnects two levels of recursion while keeping model weights fixed:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>Task-level recursion\u003C\u002Fstrong> refines the execution harness for the current task\nthrough a contrastive \u003Cstrong>Propose–Probe–Compose\u003C\u002Fstrong> step: propose edits, probe\ntheir effects on the same task, and compose a refined harness from behavioral\ncontrasts between executions, without ground-truth feedback.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Harnessing-level recursion\u003C\u002Fstrong> updates a persistent \u003Cem>improvement harness\u003C\u002Fem>\nwith adopted harnesses and evidence-supported lessons from each task to guide\nfuture refinement.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp align=\"center\">\n  \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fgoogle-research\u002Fharness2\u002Fblob\u002FHEAD\u002Fassets\u002Fharness2_pipeline.png\" rel=\"nofollow ugc noopener\">\n    \u003Cimg src=\"https:\u002F\u002Fraw.githubusercontent.com\u002Fgoogle-research\u002Fharness2\u002FHEAD\u002Fassets\u002Fharness2_pipeline.png\" alt=\"Harness² pipeline. Task-level recursion: an agent executor runs the initial harness, a harness editor proposes invariant and domain edits, candidate probes execute them, and a harness composer combines them into the final harness. Harnessing-level recursion: an experience reflector reflects on the step and updates the improvement harness used for the next task.\" width=\"100%\" \u002F>\n  \u003C\u002Fa>\n\u003C\u002Fp>\u003Cp>Across three professional benchmarks and four model–harness configurations, a\nsingle Harness² step improves GLM-5 and Gemini 3.5 Flash over their base\nharnesses, with gains of up to 11.5 points on WorkBuddy-Bench, 4.5 points on\nJobBench, and 11.8 percentage points on LAB's partial pass rate. Increasing the\ntask-level recursion budget brings further gains.\u003C\u002Fp>\n\u003Cp>The release supports \u003Cstrong>parallel and sequential recursion\u003C\u002Fstrong>, \u003Cstrong>OpenCode and\nCodex executors\u003C\u002Fstrong>, and \u003Cstrong>JobBench, LAB, and WorkBuddy-Bench\u003C\u002Fstrong>.\u003C\u002Fp>\n\u003Ch2>Video overview\u003C\u002Fh2>\n\u003Cp>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fuser-attachments\u002Fassets\u002F76b10548-b14b-4c8f-9aa5-a93662df4d56\" rel=\"nofollow ugc noopener\">https:\u002F\u002Fgithub.com\u002Fuser-attachments\u002Fassets\u002F76b10548-b14b-4c8f-9aa5-a93662df4d56\u003C\u002Fa>\u003C\u002Fp>\n\u003Ch2>Contents\u003C\u002Fh2>\n\u003Cul>\n\u003Cli>\u003Ca href=\"#video-overview\" rel=\"nofollow ugc noopener\">Video overview\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#harness-components\" rel=\"nofollow ugc noopener\">Harness components\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#install\" rel=\"nofollow ugc noopener\">Install\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#run-a-benchmark\" rel=\"nofollow ugc noopener\">Run a benchmark\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#experiments\" rel=\"nofollow ugc noopener\">Experiments\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#web-access\" rel=\"nofollow ugc noopener\">Web access\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#results-and-resume\" rel=\"nofollow ugc noopener\">Results and resume\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#citation\" rel=\"nofollow ugc noopener\">Citation\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#contributing\" rel=\"nofollow ugc noopener\">Contributing\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#license\" rel=\"nofollow ugc noopener\">License\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#disclaimer\" rel=\"nofollow ugc noopener\">Disclaimer\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2>Harness components\u003C\u002Fh2>\n\u003Cp>Harness² exposes eight editable components around the model: \u003Cstrong>system prompt,\nguardrails, memory, tool guidance, subagents, skills, scripts, and plugins\u003C\u002Fstrong>.\nThey control context construction, the tool interface, delegation, and behavior\ninside the tool loop.\u003C\u002Fp>\n\u003Cp align=\"center\">\n  \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fgoogle-research\u002Fharness2\u002Fblob\u002FHEAD\u002Fassets\u002Fcomponents.png\" rel=\"nofollow ugc noopener\">\n    \u003C\u002Fa>\u003C\u002Fp>",1793926092862]