Harness Runtime

Docs | Quick Start | Concepts | SDK | CLI | Architecture

Harness Runtime is the SDK and local runtime for repo-native harness engineering.

It helps engineering teams turn ad hoc coding-agent runs into a repeatable loop:

harvest tasks -> run in isolation -> verify -> capture artifacts -> update evals -> improve playbooks

Bring your own repository, agent, model, verifier, and workflow. Harness Runtime provides the control layer around them.

Key Message

Harness Runtime is a language-agnostic harness layer with a Python SDK.

That means:

• the package is installed with Python
• the target repository can be Python, Rust, TypeScript, Go, or a monorepo
• verification is command-based, not framework-locked
• agents and models are bring-your-own

It is not another coding-agent framework and not a hosted control plane.

Quick Start

python -m pip install -e ".[dev]"
harness init
harness harvest --from tasks
harness run task_001 --adapter shell --agent "python your_agent.py"
harness verify latest --cleanup
harness report

For benchmark, flywheel, and reference-eval flows, start here:

• Quick Start
• CLI Guide
• Benchmarks
• Harness Flywheel

Documentation

The project now has a docs structure intended for a GitHub Pages developer site:

• Docs Home
• Quick Start
• Concepts
• SDK Surface
• CLI Reference
• Benchmarks
• Architecture

Deeper design references:

• Core Abstractions
• Plugin System
• SDK Surface Notes
• Multi-Language Support Model
• Core vs Plugin Boundary

Privacy and Compliance

Harness Runtime is local-first by default:

• no code, prompts, diffs, logs, traces, or verification results are uploaded by the framework
• .harness/ contains local state and should normally stay uncommitted
• cloud usage belongs to the agent you choose to run, not to the harness itself
• provider integrations should stay explicit opt-in extensions

License

Apache-2.0