Harness Kit

Concepts

The small vocabulary that keeps agent work operational.

Harness Kit names the parts that need to be assigned, checked, logged, and improved.

Harness

The operating wrapper around agents: filesystem conventions, skills, roster, permissions, receipts, gates, and closeout rules.

Skill

A task-specific judgment contract in SKILL.md. It tells the lead agent how to run a workflow without hiding a product inside prose.

Agent

A scoped persona or provider lane used for fresh context, implementation, review, or critique.

Gate

A deterministic command that must pass before work is treated as trustworthy.

Oracle

The evidence that proves the requirement, not just a nearby green check.

Trace

A sanitized record of who ran, why, what happened, and what the lead accepted.

Roster

The configured set of external providers the lead can dispatch and compare.

Workflow

A repeatable chain such as shape -> deliver -> review -> CI -> QA -> reflect.

Provider

An external agent runtime in the roster, such as Claude, Codex, Pi, Grok, Cursor, or Antigravity.

Receipt

The structured closeout record of a provider attempt: objective, status, evidence, lead verdict, and what was accepted.

What to verify

Sources: agents, skills, harnesses/shared/AGENTS.md