Harness² is an open-source framework for refining the prompts, tools, memory, and other components around fixed AI models. It helps researchers and developers run recursive harness-improvement workflows across JobBench, LAB, and WorkBuddy-Bench.
Harness² is an Apache-2.0-licensed framework from the google-research repository for recursive agent harnessing in open-world tasks. It focuses on improving the system around a fixed model, including prompts, tools, memory, skills, orchestration, and other editable behavior rather than changing model weights.
Its task-level workflow uses Propose-Probe-Compose: it tests candidate harness edits on the current task and uses behavioral evidence from those executions to construct a refinement. A second, persistent level records adopted harnesses and evidence-supported lessons so later tasks can draw on prior refinement work.
The framework treats the surrounding agent system as editable components: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins. These components govern context construction, tools, delegation, and behavior in the tool loop.
For a current task, Harness² proposes harness edits, probes their effects on that same task, and composes a refined harness from contrasts in execution behavior. The documented workflow does not require ground-truth feedback for this step.
After a task, the system retains adopted harnesses and evidence-supported lessons in a persistent improvement harness intended to guide later refinements.
Parallel recursion explores multiple general and domain-specific candidates, while sequential recursion carries a refined harness and accumulated evidence through successive steps. The `--k` setting controls the task-level recursion budget.
The release supports OpenCode by default and can use Codex with `--substrate cx`. It includes workflows for JobBench, LAB, and WorkBuddy-Bench.
Command-line workflows cover setup, environment checks, refinement, judging, and reporting. Results retain task summaries, deliverables, adopted harnesses, reflection notes, and evaluation outputs.
Run controlled refinement experiments against professional-work tasks in JobBench, using the benchmark setup, runner, judging, and report commands to compare configured harness behavior.
Use the LAB integration for legal-work benchmark tasks, where the documented setup includes the Pandoc and Poppler dependencies required by that benchmark.
Run WorkBuddy-Bench subsets for office, web, or code tasks. The documented workflow requires Docker and carries the selected subset through commands after setup.
Explore candidate harness changes broadly with parallel recursion, or iteratively deepen a single refinement path with sequential recursion. This suits experiments where the choice between breadth and cumulative refinement matters.
Start with the local demo to exercise both recursion modes using deterministic editors and real Python executions before configuring model-backed benchmark runs.
For benchmark runs, the project specifies Linux and Python 3.12. Its Python package also supports Python 3.13. The documented setup additionally calls for uv, Node.js 20 or later, and the Google Cloud CLI; some benchmarks add their own dependencies.
The repository documents JobBench, LAB, and WorkBuddy-Bench. A benchmark is installed with `harness2-setup`, then checked with `harness2-check` before a refinement run.
Use `harness2` to run refinement, followed by `harness2-judge` and `harness2-report` for evaluation and reporting. `--dry-run` previews task selection without model calls. Judging and reporting need the same benchmark, executor, task subset, recursion budget, mode, and seed as the refinement run.
Results are saved under `runs/`, organized by benchmark and, where applicable, subset. They include task summaries, deliverables, adopted harnesses, reflection notes, and evaluation results. Repeating the same command resumes a run; changing models or task selection calls for a new `HARNESS2_RUNS_DIR`.
CreateOS Sandbox is an isolated compute environment for running code and agent workloads inside Firecracker micro-VMs. It is designed for workflows that need machine-level isolation, private networking between sandboxes, and programmatic control through SDK, CLI, or MCP.
GitBot is a local browser-based workspace for creating reusable coding bots powered by Claude Code, Codex, or OpenCode. It helps developers run those agents across repository folders with persistent threads and configurable permissions.
AakarDev AI helps teams manage AI provider access, project-level setups, logs, and analytics from one dashboard. It supports BYOK workflows and lists providers including OpenAI, Google Gemini, Anthropic, Groq, Mistral AI, and Perplexity AI.
Trigger.dev chat agent is a durable AI chat backend for developers building stateful conversations that can survive refreshes, crashes, and long-running turns. It connects with the AI SDK `useChat` flow and runs on managed infrastructure with no timeout on a turn.
ByteAsk is a terminal-first AI coding agent for C and C++ that edits repositories and verifies changes with the real compiler, debugger, sanitizers, and tests before showing a diff. It offers a free tier plus paid plans, with editor connectors and zero-retention handling described in the source.
Epherem is a static security scanner for web apps built with AI coding tools such as Lovable, Bolt, v0, Replit, Cursor, and Claude Code. It checks source files and dependencies for common security mistakes, then provides categorized findings and suggested fixes without running the app.
