What Is an AI Harness?

#ai #harness #architecture

What Is an AI Harness?

Key jargon

Term	Plain-language meaning
Harness	The application layer that prepares model calls and controls what happens around them.
Orchestration	Coordinating steps, models, tools, state, and recovery.
Guardrail	A technical or procedural control that restricts unsafe or invalid behavior.
Run record	A durable account of inputs, decisions, tool calls, outputs, and outcomes.

Key concepts

The model proposes language or actions; the harness owns execution and control.
A production harness turns a probabilistic component into an observable, bounded system.

Concept map

flowchart LR
    A["User goal"] --> B["Harness assembles and controls"]
    B --> C["Model proposes"]
    C --> D["Harness validates and delivers"]

A raw model accepts input and returns output. A harness decides what input it sees, what tools it may call, what state survives, what output is accepted, and what happens next.

flowchart LR
    U[User or event] --> P[Policy and task contract]
    P --> C[Context assembler]
    C --> M[Model adapter]
    M --> V[Validator]
    M <--> T[Tool gateway]
    C <--> S[State and memory]
    V --> H[Human approval or output]
    P --> O[Observability and evals]
    T --> O
    V --> O

Harness responsibilities

Authenticate users and bind tenant/project identity.
Build trusted and untrusted context with provenance.
Route tasks to a model/version and enforce budgets.
Describe and broker tools with least privilege.
Validate structured outputs and external effects.
Persist state, checkpoints, approvals, and evidence.
Recover safely from timeouts and ambiguous outcomes.
Measure quality, latency, cost, safety, and drift.

Exercise

Draw the harness around a familiar chat assistant. Mark which components you can observe and which are assumptions.