cobusgreyling

AI Harness Engineering — AI skill for Claude Code

AI community

Minimal Agent Harness — Six harness components demonstrated with NVIDIA Nemotron 3 Super.

How to install AI Harness Engineering

This entry records only its repository, not the path inside it, so there is no exact command to give. Open cobusgreyling/ai_harness_engineering and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What AI Harness Engineering does

Minimal Agent Harness — Six harness components demonstrated with NVIDIA Nemotron 3 Super.

Alternatives in AI

CoreCoder — Minimal AI coding agent (~1,400 LoC Python) inspired by Claude Code 616 ★
Maestro Companion — Quick execution for small tasks — minimal run lifecycle (start + done) with evidence recording 530 ★
Askimo — AI Client for chat, RAG, Skills, MCP tools, and agents 470 ★

README

AI Harness Engineering

![Agent Harness Engineering](images/header.svg)

A **harness** is the software system that governs how an AI Agent operates. It manages tools, memory, retries, context engineering and verification so the model can focus on reasoning.

This repo is a complete, working implementation of all six harness components — plus advanced features like guardrails, budget enforcement, sub-agents, and persistent memory — both as a CLI demo and a Gradio GUI, powered by **NVIDIA Nemotron Super 49B**.

![Architecture Stack](images/03-architecture-stack.png)

What's Inside

File	Description
`harness_core.py`	Shared module — all components, tools, configs, and the main `Harness` class
`harness-demo.py`	CLI agent with 9 tools, retry loop, token tracking, YAML configs, and comparison mode
`harness-gui.py`	Gradio GUI with five tabs: Run Task, Compare Configs, Audit Log, Observability, Replay
`configs/`	YAML config presets (autonomous, conservative, minimal, guardrails, budget-limited)
`harness-engineering-blog.md`	The full blog post on harness engineering

The Six Harness Components

![Six Components](images/05-six-components.png)

#	Component	Class	Role
1	Tool Integration	`ToolRegistry`	Register and execute 9 tools with risk levels
2	Memory & State	`MemoryManager`	Key-value working memory with optional SQLite persistence
3	Context Engineering	`ContextEngine`	Dynamically assembles system prompt + tool descriptions + memory + retry feedback + guardrail instructions
4	Planning	`Planner`	LLM call to decompose a complex task into discrete steps
5	Verification	`Verifier` + `Guardrails`	Rule-based checks + PII detection, toxicity filtering, code injection prevention
6	Modularity	`HarnessConfig`	Toggle any component ON/OFF, load from YAML presets