Four Layers, Four Responsibilities

Layer	Owns	Typical artifacts	Selection question
Model	Generation and reasoning	Weights, API, context window	Which model meets the task, latency, privacy, and cost constraints?
Repository harness	Project-specific instructions and controls	AGENTS.md, CLAUDE.md, skills, hooks, policies	What behavior must remain portable with the repository?
Runtime harness	Tool loop, permissions, state, recovery	CLI, IDE agent, desktop or cloud runtime	Who owns plan, act, observe, repeat?
Orchestrator	Queues, workspaces, budgets, multiple runs	Scheduler, fleet manager, task board	What must coordinate more than one runtime or agent?

Frameworks, SDKs, sandboxes, memory systems, evaluation tools, observability platforms, and protocols sit beside or below these layers. LangGraph can help build a runtime; E2B can isolate its execution; Mem0 can persist memory; Langfuse can observe it; MCP can connect tools. None of those roles alone proves ownership of the coding loop.

The term meta-harness has two incompatible uses. Optimizer research uses it for a system that changes one or more harness layers under evaluation. Products such as Omnigent use it for a common interface that dispatches tasks to existing harnesses. Keep the roles separate: this guide classifies a dispatcher as an orchestrator or control plane, while a harness optimizer changes the system that will perform future runs. The generated directory below preserves each pinned source's wording, so Omnigent's row retains its upstream meta-harness label even though the guide layer is orchestration. See the Databricks cost-management resource evaluation for the terminology boundary.