The 6 Levels of AI Engineering

Short answer

The six levels of AI engineering are basic chat, prompt engineering, context engineering, loop engineering, graph engineering, and harness engineering.

Basic chat is a one-off answer. Prompt engineering improves the instruction. Context engineering improves what AI can see. Loop engineering improves how AI works through steps, checks output, retries, and stops. Graph engineering improves how knowledge, tools, states, decisions, and workflows connect. Harness engineering improves the operating layer around the agent.

The 6 levels

Use this ladder when deciding what to fix first.

How it matches the AI skill ladder

The seven-level AI skill ladder is about the person using AI. This six-level engineering ladder is about the system being built.

They fit together cleanly. Beginners start with basic chat and better prompts. Power users build context. Operators design workflows, loops, graphs, and runtime controls. The mistake is jumping to the advanced layers before the simpler ones work.

A simple coding-agent example

Imagine asking an AI coding agent to fix a bug.

At the prompt level, you tell it the bug, the expected behavior, and the output you want. At the context level, you give it the relevant files, error message, screenshots, tests, and product rule. At the loop level, it reads the code, makes a small change, runs the test, checks the result, and tries again if it failed.

Graph engineering appears when that work is no longer one straight path. The agent might branch into frontend, backend, and database checks. It might pause for human approval before changing auth. It might route a risky migration to review while letting a copy fix ship. Harness engineering is the environment that makes all of that possible: tools, permissions, logs, memory, sandboxes, and handoffs.

Basic chat

Basic chat is the starting level. You ask AI a question and get an answer: a summary, explanation, brainstorm, rewrite, comparison, or checklist.

It is useful, but it is not yet engineering. There is no reusable instruction, no durable context, no verification loop, no state, and no runtime. That is why basic chat is good for learning and weak for repeatable business work.

Prompt engineering

Prompt engineering is still useful. It is the practice of giving clear instructions, constraints, examples, and desired output formats.

Use it when the model already has enough information but gives the wrong shape of answer. If the answer is too generic, too long, too vague, or formatted badly, the prompt may be the problem.

Sources: OpenAI: prompt engineering guide

Context engineering

Context engineering is the next layer. It is the work of giving the model the right information at the right time: source files, examples, business rules, retrieved knowledge, memory, tools, and constraints.

LangChain describes context engineering for agents as managing what enters the model context through writing, selecting, compressing, and isolating context. In plain English: do not dump everything into AI. Give it the right material for the current step.

Sources: LangChain: context engineering in agents

Loop engineering

Loop engineering is what happens when AI needs to keep working, not only answer once. A loop gives the model a goal, context, tools, checks, retry rules, and a stop condition.

This matters for coding agents, content workflows, lead follow-up, research, audits, and any task where the first answer is not enough. A good loop does not run forever. It knows when to verify, when to retry, when to ask for help, and when to stop.

Graph engineering

Graph engineering is the connection layer. It becomes useful when the hard part is not one answer or one loop, but the connections between entities, documents, sources, tools, states, branches, approvals, and decisions.

There are two common meanings. GraphRAG uses knowledge graphs to retrieve connected evidence before answering. Agent graphs and state machines use graph-shaped control flow to decide what an AI system is allowed to do next. They are related, but they solve different problems.

Sources: Microsoft GraphRAG documentation, LangGraph overview

Harness engineering

Harness engineering is the runtime layer around the model. It includes the system prompt, tool definitions, memory, file access, browser or shell access, model routing, middleware, permissions, logs, traces, sandboxes, approval gates, and handoffs.

This is why two teams can use the same model and get very different results. One team gives the agent clean tools, narrow permissions, durable state, and observable execution. The other gives it a vague prompt and hopes the model behaves. The model may be similar; the working conditions are not.

A good harness does not make the agent magically intelligent. It makes the agent safer, easier to resume, easier to inspect, and less likely to waste tokens by rereading the same messy context every turn.

Which one do you need?

Pick the layer that matches the failure.

How the layers work together

A serious agent usually needs several levels, but not all at the same depth.

Imagine a research-and-publishing agent. Basic chat helps explore the idea. The prompt defines the task. The context provides sources, brand rules, and examples. The loop drafts, checks, revises, and stops when the evidence passes. The graph controls branching: research, draft, review, approval, publish, or retry. The runtime supplies tools, memory, sandboxing, logs, permissions, and handoffs.

That is the clean sequence: basic chat, prompt, context, loop, graph, harness. Each level answers a different failure.

Reference links

These sources anchor the AI engineering level map.

Sources: OpenAI: prompt engineering guide, LangChain: context engineering in agents, OpenAI Agents SDK, LangChain: agents overview, Microsoft GraphRAG documentation, LangGraph overview, Microsoft AutoGen: GraphFlow workflows

Read this cluster in order

These ideas are not competing terms. They are levels. Start with the full map, then move through instruction, context, repeated work, control flow, connected knowledge, runtime, token management, and the human skill ladder.

Use the sequence below when deciding what to fix first. If the work only needs a quick answer, basic chat is enough. If the instruction is unclear, improve the prompt. If the model lacks facts, improve context. If the work needs checking and retries, design a loop. If the work branches, use a graph. If the agent needs tools and permissions, improve the harness.

The human roles behind the levels

The engineering levels explain the system. The human roles explain the work.

A Prototyper works closest to basic chat and prompt engineering. A Builder moves into context and harness. A Sweeper improves loops and security. A Grower connects content, distribution, and measurement. A Maintainer keeps the system useful after launch.

Sources: The 5 AI Roles That Matter More Than Job Titles

Final answer

Basic chat is the answer level. Prompt engineering is the instruction level. Context engineering is the information level. Loop engineering is the repeated-work level. Graph engineering is the connection and control-flow level. Harness engineering is the runtime level around the model.

Learn them in that order. It keeps AI from feeling like magic and turns it into a system you can improve one layer at a time.