AI Harness Reference Architecture

Key jargon

Term	Plain-language meaning
Control plane	Components that decide policy, routing, permissions, and workflow state.
Data plane	Components that carry prompts, retrieved context, tool data, and model outputs.
Adapter	A boundary translating a stable internal contract to a provider-specific interface.
Policy engine	Code that evaluates whether an operation is allowed.

Key concepts

Separate control decisions from model-generated content so text cannot grant itself authority.
Stable internal contracts isolate the application from changing providers and frameworks.

Concept map

flowchart LR
    A["Request enters control plane"] --> B["Assemble governed context"]
    B --> C["Invoke model and tools"]
    C --> D["Record evaluate return"]

Layer	Responsibility	Evidence
Interface	User intent, authentication, confirmation, accessibility	Request and user/session ID
Policy	Scope, data rules, tool rules, budgets, stop gates	Policy version and decision
Orchestration	Workflow/graph, transitions, queues, retries	Run and step state
Context	Instructions, history, retrieval, source trust	Context manifest and hashes
Model	Provider adapter, version, parameters, routing	Model ID and usage
Tool gateway	Schema, authorization, execution, result normalization	Call, identity, response, effect ID
Memory/state	Durable facts, preferences, artifacts, checkpoints	Provenance and retention metadata
Validation	Schema, business rules, citations, tests, safety checks	Assertions and disposition
Operations	Tracing, metrics, cost, incidents, evals	Trace and scorecard

Design rule

Each layer should be replaceable and testable. If changing the model requires rewriting permissions, storage, and business logic, boundaries are too coupled.

Exercise

Map an existing RAG chatbot into the table. If a layer is absent, write “absent” rather than assigning it to the model.