Choose theme
AI Resources
Harness²
Harness² is a Google Research framework for experimenting with the system around an AI model: its prompts, tools, memory, skills and orchestration. It refines that harness while keeping model weights fixed.
The repository includes a local demo, benchmark adapters and saved experiment outputs. It is research code, and the README says it is not an officially supported Google product. Use this as a first read, not a recommendation. Open the original project before trusting details like terms, limits, privacy, cost, setup, or safety.
What it is
Experiments around a fixed model
A task-level loop proposes harness changes, probes them on the same task and composes a refined harness from differences between executions. This refinement does not use ground-truth feedback.
Why it stands out
Refinement within and across tasks
After adopting a harness, a second loop saves it and evidence-supported lessons for later tasks in the same domain. Task-level refinement and the persistent improvement library therefore work at different timescales.
Availability
Source, demo and benchmark runners
The public repository provides Python code, tests and setup guides. The project declares Apache-2.0; see the repository and each upstream benchmark for current terms.
Why it matters
What makes it useful
Agent behavior depends on more than the model checkpoint. Harness² exposes editable components including prompts, tool guidance, memory and skills, making those parts available for experiments. Its local demo runs the recursion and library workflow with deterministic editors and real Python execution, without model credentials. That gives a builder a smaller way to examine the workflow before configuring model-backed benchmarks.
What to know
Where it fits
This is a framework for running agent experiments. The benchmark adapters cover JobBench professional tasks, LAB document-based legal-work tasks and WorkBuddy-Bench office, web and code tasks. Refinement, judging and reporting are separate commands. Benchmark support describes an evaluation workflow; it does not establish suitability for doing a person's professional work.
Notable points
What stands out
Parallel recursion probes a pool of candidate harnesses and uses their execution evidence in compositions. Sequential recursion carries a refined harness and its accumulated evidence through successive steps. Both save task artifacts, adopted harnesses and reflection notes. Failed evaluations remain unscored, so a missing judgment is different from a measured zero.
Before using
What to review
The setup guide recommends Linux and Python 3.12; package metadata also supports Python 3.13. Model-backed runs need additional runtimes, benchmark dependencies and model access. They incur API costs; the experiment guide recommends a small task selection first.
The repository says executor network access is unrestricted and Codex runs with its sandbox disabled. Its evaluation guide recommends network-level restrictions to reduce access to public benchmark answers and grading criteria. Review the documented execution access before running experiments.
Keep benchmark, executor, subset, recursion budget, mode and seed aligned across refinement, judging and reporting. The seed controls task ordering, not model sampling. Setup checks do not establish that every configured model is available.
Reader fit
Who may find it relevant
Agent researchers and developers comparing harness designs are the clearest audience. The credential-free demo offers an entry point for understanding the mechanism. Running the model-backed studies calls for comfort with command-line setup, model credentials and benchmark-specific dependencies.
Editorial note
Why LifeHubber lists it
LifeHubber lists Harness² because its proposed improvements leave concrete material to examine: the adopted agent setup alongside execution artifacts and reflection notes. For a builder investigating how a harness changed its approach to a task, those records connect the chosen instructions and tools with the work produced. They provide a basis for reviewing a proposed design change, while judging the quality of that change still needs a separate evaluation.
Source links
Source materials
Reader note
Before relying on this entry
LifeHubber lists entries to help readers inspect AI projects, not to endorse them or prove they are safe, suitable, accurate, maintained, or right for a specific use. We do not verify every entry in depth. Before relying on anything listed, review the original materials, terms, privacy practices, limits, and risks that matter for your situation.
More in AI Agents
Keep browsing this category
Explore more AI agent projects.
Paperclip
paperclipai/paperclip
A self-hosted server and dashboard for coordinating agent teams through companies, goals, roles, issues, heartbeats, budgets, approvals, and persistent activity records.
Anatomy
wheresryan22/anatomy
A coding-agent skill and drawing kit for interactive technical illustrations, with editable SVG and web output, shared motion and readout models, browser checks and optional beta 3D.
Open Multi-Agent
open-multi-agent/open-multi-agent
A TypeScript-native framework for coordinating multi-agent work from a goal into a task DAG, with auto task decomposition, parallel execution, plan preview and replay, human approval hooks, MCP tools, provider routing, observability, and sandboxed filesystem defaults.
For project maintainers
Listed here? You can use the badge.
If you maintain a project with a current LifeHubber listing, you may add the optional “Listed on LifeHubber AI Resources” badge to its README, docs, or website. No introduction or permission request is needed.
