The Harness Equation

The overall architecture of OpenHarness can be summarized in one simple equation: Harness = Tools + Knowledge + Observation + Action + Permissions Let's break down each part:

Tools are the interfaces through which an Agent interacts with the outside world. OpenHarness has 43+ built-in tools covering: • File operations: Read, Write, Edit, Glob, Grep • Shell commands: Bash • Search: WebFetch, WebSearch • Agent collaboration: Agent, SendMessage • Task management: TaskCreate, TaskUpdate • MCP integration: MCPTool • Workflow control: EnterPlanMode, EnterWorktree Every tool follows a unified interface: receive parameters → execute operation → return result.

An Agent can't fit all knowledge into a single prompt. OpenHarness's Skills system allows loading domain-specific knowledge on demand, like a library — grab the book you need when you need it.

Every time a tool finishes executing, the result is fed back to the LLM. This is "observation" — the Agent can see what effect its action had, then decide what to do next.

An Agent doesn't complete a task in one shot. It continuously loops (Agent Loop): think → call a tool → observe the result → think again. This loop is the Agent's heartbeat — we'll cover it in detail in Chapter 4.

Letting AI operate freely is dangerous. OpenHarness provides three permission tiers to ensure the Agent doesn't perform inappropriate actions. This is one of the most important design elements of the Harness.

At the code level, OpenHarness breaks those five capabilities into 10 subsystems: 1. Engine — Agent Loop core cycle 2. Tools — 43+ built-in tools 3. Skills — On-demand knowledge loading 4. Plugins — Extension mechanism 5. Permissions — Safety checks 6. Hooks — Lifecycle events 7. Commands — 54 user-interaction commands 8. MCP — Model Context Protocol integration 9. Memory — Cross-session persistence 10. Tasks — Background task management In the following chapters, we'll dive deep into each subsystem.