FoxTrove Field Guide

AI Operating Layers

The AI market is full of overlapping words: agents, workflows, skills, loops, harnesses, memory, connectors, evals, and orchestration. The words matter less than the structure underneath them.

AI becomes business leverage when a model is connected to the right context, given a repeatable process, allowed to use the right tools, checked by evidence, and embedded into a workflow with human control.

Layer Model

From raw AI to an operating system.

Layer 0

Raw AI

A model can read, reason, and generate. Useful, but disconnected from how your business actually works.

Layer 1

Prompt

The instruction someone gives the model. Better prompts help, but the value still depends on the person typing.

Layer 2

Context

The documents, examples, data, history, and constraints the model needs to answer correctly.

Layer 3

Connector

A safe link into business systems: CRM, email, databases, files, browser, ERP, or APIs.

Layer 4

Skill / Playbook

Reusable process knowledge. The company way of doing work, written down for AI to follow.

Layer 5

Workflow

A defined sequence of steps. This is where AI starts becoming process, not just chat.

Layer 6

Agent

An AI system that can choose steps and use tools inside defined boundaries.

Layer 7

Harness

The control layer: permissions, tests, logs, sandboxing, state, approval gates, and rollback.

Layer 8

Loop

A recurring system that finds work, does a first pass, verifies, records state, and escalates exceptions.

Layer 9

AI Operating System

The business layer: ownership, governance, ROI reporting, cost control, and continuous improvement.

Plain English

Use the fewest words that preserve the distinction.

The mature buyer does not need a lecture on agent frameworks. They need to know which layer creates value, which layer creates risk, and who owns the result.

Workflow

The business process being improved. This is the unit FoxTrove diagnoses first.

Automation

A workflow that runs on a trigger, schedule, or event.

Agent

Use this word only when AI chooses next steps or tools, not for every chatbot.

Harness

The controlled environment that keeps an agent observable, permissioned, testable, and reversible.

Loop

A workflow or agent that repeats with state, verification, and escalation.

Human Gate

The approval point before external, financial, legal, destructive, or high-risk action.

Proof Loop

Evidence, verification, state, ROI review, and improvement.

AI You Own

The local or hybrid runtime for sensitive, high-volume, or cost-sensitive AI work.

Decision Rules

Not every workflow deserves an agent.

The point is not maximum autonomy. The point is the right amount of autonomy for the workflow, with the right controls around it.

If

If the steps are predictable

Use workflow automation before an agent.

If

If the work varies but has clear boundaries

Use a scoped agent with tools and fallback paths.

If

If the work recurs every day or week

Use a loop with durable memory and a stop condition.

If

If the system touches customers, money, legal, or production

Add a human gate.

If

If data sensitivity or token cost is material

Route the work through AI You Own or a hybrid runtime.

If

If quality cannot be tested

Do not automate yet. Build the evidence and eval layer first.

FoxTrove Position

We do not sell agents as the answer to everything.

We identify where AI should assist, where it should automate, where it should act as an agent, and where it should stay behind a human gate. Then we build the operating layer around it: context, tools, workflows, memory, verification, governance, and ROI.

See the Operating Blueprint →