← Docs · The Agentic Harness
How the harness works
A harness gives people and AI agents a shared way to work: reusable knowledge, tools to act with and checks that keep a repository consistent. Understanding its parts lets you choose what to activate and what evidence a passing audit actually provides.
What it is
An agentic harness is the working environment around an agent. The KI Agentic Harness has four source shelves: skills, specialist agent roles, evaluations and runtime hooks. Its installed payload contains skills, agent roles and hooks; evaluations stay with the source. MCP servers live in separate repositories, with a portable inventory describing the connections a runtime should use.
A skill supplies a reusable capability. Some govern an artefact, such as writing or repository structure; others guide a process, such as planning or reviewing work. What a skill is explains that distinction, and the catalogue lists the available capabilities.
How it works
The harness serves both human and LLM agents. Work that a deterministic operation can decide belongs in a mechanical check. Questions that require interpretation remain judgement for a person or an AI agent.
The mechanical baseline runs without a model. The installed ki CLI reads a repository's declarations, resolves compatible installed harnesses and executes their registered rubrics. It reports which judgement criteria remain unassessed, so a clean mechanical result does not silently claim that those questions were reviewed.
The harness publishes capability definitions and their safe repairs. The CLI owns installation, activation, execution and reporting. A nearby checkout is used only through an explicitly selected development binding.
How you use it
Start by installing a harness on your machine. Then declare the skills that govern a repository and ask the agent for the outcome you want.
For example, a fictional Harbour Atlas island can declare its writing standard, run a read-only audit and ask an agent to review whether its explanations help a newcomer. The audit checks the mechanical rules; the review supplies the judgement.
What it provides
The same declared standards can support an interactive agent session, a human running commands and automated checks in CI. Shared rules stay with their owning skills, while repository-specific choices remain in tracked declarations and guidance.
AGENTS.md carries portable repository orientation. A runtime-specific file such as CLAUDE.md holds only the detail that belongs to that runtime. Agent environment explains the separate machine and runtime layers.
Boundaries
Installing a harness makes its capabilities available; activation selects which ones apply. A skill declaration does not prove that a running session has loaded it or that a connected service is healthy.
A passing audit covers the declared checks it ran. It does not grant permission to publish, change credentials or perform other actions outside the task's authority.