Agentic AI · Service Design

What is an
AI harness?

The behaviour of an AI agent is decided mostly by its harness, not by the model inside it. A plain-language explainer for business and service designers — and a question about who should be designing it.

Read the explainer Skip to the design angle

New vocabulary from the agentic AI deep dive

Part 1

What is an AI “harness”?

Why every AI agent is really two things: a model, and everything built around it.

People say things like “the agent harness matters as much as the model” or “we improved the harness, not the model”. The word comes from classic engineering. A horse harness turns raw animal power into something that can pull a cart in a chosen direction. A wiring harness in a car bundles hundreds of loose cables into one organised system connecting the engine to everything else. In both cases the harness does not create the power — it directs it, connects it, and keeps it under control.

In AI the idea is the same. A model on its own is like an engine on a workshop table: impressive, but going nowhere. All it can do is receive text and produce text. It cannot open a file, check a calendar, browse a website or update a customer record. The harness is the software built around the model that gives it those abilities and decides which abilities it gets.

In one sentence: a harness is the surrounding structure of instructions, tools, memory and guardrails that lets an AI model act in the real world, and keeps that action safe and on track.
The model alone AI model (the “engine”) Very smart, but it can only read text in and write text out. It cannot touch anything. add a harness The same model in a harness The harness AI model same engine Instructions its job & rules Tools hands to act with Memory what it knows Guardrails limits & approvals Now it can do useful work in the real world — safely.
Fig. 1 — The model alone vs. the same model wrapped in a harness.

Part 2

From answering to acting: the agentic loop

This is where “agentic AI” enters. A chatbot answers one question at a time. An agent is given a goal and works towards it in steps: it thinks about what to do, does it, looks at what happened, and decides what to do next. The model does the thinking in that loop. The harness runs everything else — it executes the actions, feeds results back to the model, tracks progress, and stops the loop when the goal is reached or something goes wrong.

the model thinks the harness acts Goal “Book my trip” 1 · Think model decides next step 2 · Act harness uses a tool 3 · Observe harness returns the result 4 · Check done, or loop again? not done: loop Done: deliver the result “Your trip is booked” Example tools · search the web · read a file · check a calendar · send an e-mail · update a system
Fig. 2 — The loop the harness runs, over and over, until the goal is reached.

What is inside a harness

Instructions

What the model’s job is, what tone to use, which rules to follow.

Tools

The actions it may take — search the web, read documents, update a system.

Memory

What the harness shows the model at each step: the goal, the conversation so far, relevant documents, results of earlier actions.

Guardrails

The limits: which actions need human approval, what the agent may never do, when it must stop and hand to a person.

Part 3

Why the harness matters so much

The same model behaves very differently depending on the harness around it. Give it web search and a friendly tone and you have a chat assistant. Give it a code base and permission to run programs and you have a coding agent. Give it customer systems and strict policies and you have a customer service agent. When companies release a better “agent” without a new model, the improvement usually happened in the harness.

Same AI model Chat assistant harness Tools: web search · Memory: this conversation · Guardrails: answers only, asks before acting → a helpful conversation partner Coding agent harness Tools: read/write files, run code · Memory: your whole project · Guardrails: asks before big changes → a software developer that works for hours Customer service harness Tools: customer systems, billing · Memory: the customer’s history · Guardrails: strict company policy → an agent that resolves customer requests
Fig. 3 — One model, three harnesses, three very different agents.

Part 4

What to remember

  1. 01

    The model is the engine; the harness is everything that turns it into a useful, safe agent.

  2. 02

    Agentic AI = a model running in a loop, with the harness executing actions and feeding back results.

  3. 03

    Instructions, tools, memory and guardrails are design choices. Designing them well is where much of the real product work happens.

  4. 04

    When you evaluate an AI agent, ask about the harness, not only the model inside.

For business & service designers

Where the harness ends,
the experience begins

As business and service designers, we oversee the customer experience holistically. We shape how customers move across touchpoints — digital, assisted and automated — and we make sure what gets built is desirable, feasible, viable and sustainable. Journey maps, service blueprints, value propositions: our craft is deciding how a service should behave before anyone builds it.

Right now, harness engineering is treated as a purely technical challenge. But look again at what a harness contains. Every one of its four parts is an experience decision wearing a technical costume.

Instructions → tone of voice & service scripts
What the agent says, how it says it, which rules it follows. Deciding this has always been tone-of-voice and conversation-design work — not a configuration file.
Tools → what the service can actually do
Which systems the agent may act on determines what it can resolve for a customer in the moment — the difference between an agent that helps and one that apologises.
Memory → continuity of the journey
What the agent remembers about the customer between steps is what makes a journey feel continuous instead of restarting at every touchpoint.
Guardrails → moments of truth
When the agent must stop, refuse, or hand over to a person. The human handoff is the new line of visibility — and designing it is classic service design.

What will we do in the future?

The honest answer: much of what we already do, applied to a new material. Four places where our role grows rather than shrinks:

  1. Blueprint agent behaviour

    Service blueprints gain a new layer. Alongside frontstage and backstage we will map the agentic layer: which steps an agent thinks through, acts on, and hands back to people or systems.

  2. Design the handoffs

    Every guardrail is a service moment. We will define where a human must take over — escalation, approval, refusal — and make those transitions feel deliberate instead of abrupt.

  3. Write the quality bar

    The evaluation harness runs an agent through hundreds of scenarios before launch. Someone has to decide what “good” looks like in each of them. That is customer-experience criteria work, and it belongs to us.

  4. Sit at the harness table

    Instructions, tools, memory and guardrails will be decided somewhere, by someone. If designers are not in that room, the customer experience is being designed by default rather than by intent.

The question to take with you

Where does the harness end,
and where does the experience you own begin?

Next topic

What are evals?

How do you know an AI agent is any good — and who decides what “good” means? Why the quality bar is an experience decision.

Read the explainer →