• The Prompt Is Not the Product, illustrating the five layers required to build practical AI agents.

What prompt, context, harness, loop and graph engineering mean—and why the system around AI creates more value than the model itself.

AI agents for business leaders are no longer simply better chatbots. They can pursue goals, use tools and complete work—but only when the right context, controls and human oversight surround them.

Executive summary

AI is moving from answering questions to completing work. That is the real significance of AI agents.

An agent can pursue a goal, use tools and complete several steps without waiting for a new instruction each time. But the model is only one part of the system. Reliable agents require five layers:

  1. Prompt engineering: defining the task.
  2. Context engineering: supplying the right knowledge and business rules.
  3. Harness engineering: providing tools, permissions, controls and an audit trail.
  4. Loop engineering: enabling the agent to act, check, adjust and continue.
  5. Graph engineering: coordinating multiple agents, systems and people.

These are not competing trends. They are nested layers. A prompt sits inside context. Context sits inside a harness. The harness enables a loop. Multiple loops connect to form a graph.

The challenge is not simply to buy more AI. It is to redesign work while retaining human intent, judgement and accountability.

Why AI Agents for Business Leaders Need More Than Prompts

After reviewing more than 1,000 AI projects and speaking with hundreds of AI leaders, I keep seeing the same mistake.

Organisations focus on the model while underinvesting in the system around it. They write a better prompt, buy another tool and launch another pilot. But a prompt is not a process—and a model is not an agent.

Imagine the model as a highly capable new employee. A prompt is the task you give them. It does not provide the customer record, company policy, spending authority, quality standard or escalation route. Without those elements, even a brilliant employee will be inconsistent.

This helps explain the gap between adoption and impact. McKinsey found that more than 80% of respondents were not seeing tangible impact on enterprise-level earnings from generative AI. BCG reported that only 26% of companies had developed the capabilities needed to move beyond pilots and generate value.

The problem is not always the intelligence of the model. It is often the design of the work around it.

The five layers of practical AI agents

Layer Business translation Leadership question
Prompt The task brief What do we want the AI to do?
Context The briefing pack What must it know?
Harness The operating controls What can it access and change?
Loop The working rhythm How does it check, improve and stop?
Graph The team structure How do agents, systems and people coordinate?

The easiest way to understand these layers is to follow one customer complaint through the system.

1. Prompt engineering: writing the task brief

Prompt engineering means giving AI a clear instruction: the goal, constraints, desired format and quality standard.

For example:

Review this customer complaint, identify the underlying issue, draft a response in our tone and highlight anything requiring a refund decision.

This helps with one-off work. But a prompt cannot retrieve information it cannot access, enforce a permission or confirm that an action succeeded.

2. Context engineering: assembling the briefing pack

The complaint cannot be resolved properly without the order history, delivery status, previous conversations, refund policy and current stock level.

Context engineering decides what the agent receives at each stage—and what it should not see. Irrelevant or outdated material can distract the model and increase cost.

Anthropic describes context as a finite resource. In business language, context engineering gives the agent the smallest useful briefing pack for the decision in front of it.

As models become widely available, advantage will come from context they do not have: how your customers behave, how operations really work and which exceptions matter.

Context should be treated as infrastructure, not as a folder of documents added at the end.

3. Harness engineering: building the safe operating environment

The harness is everything around the model that enables reliable work: tools, data, permissions, failure handling, policies, approvals and audit records.

If the model is the engine, the harness is the vehicle: the steering, brakes, dashboard, seat belts and rules about who is allowed to drive.

Our agent may read the order, check delivery and issue a credit of up to €25. Anything larger requires approval, and bank details remain completely blocked. The harness also verifies that the record was updated.

The harness turns a probable answer into a controlled process. OpenAI’s work on harness engineering highlights tools, measurements, logs and verifiable outcomes.

4. Loop engineering: designing how the agent completes the work

A traditional prompt produces an answer and stops. An agent operates in a loop:

Goal → act → observe → check → adjust → stop or continue.

The agent proposes a resolution, records the action and verifies completion. If something fails, it tries an approved recovery step or escalates.

Loop engineering designs that cycle. IBM describes workflows in which agents act, observe, decide and iterate towards a goal with minimal human prompting.

The most important part is not endless autonomy, but a test of success. What proves completion? When must the agent stop? How many attempts are allowed?

Poor loops repeat mistakes at machine speed. Good loops combine autonomy with verification. I explore this further in The Move to Loop Engineering.

5. Graph engineering: coordinating the whole system of work

Graph engineering is an emerging term. In practical terms, it means designing the network through which work moves.

A node might be a specialist agent, rules-based check, database, human approval or customer interaction. The connections determine what happens next and which decisions redirect the work.

A complex complaint might activate one agent to review the conversation, another to inspect the account and a third to check policy. A manager approves significant compensation. Each component has a defined role; none has unlimited authority.

Rather than asking one agent to do everything, graph engineering coordinates several loops and decision points across an end-to-end outcome.

The simplest way to remember the difference is:

A prompt controls one answer. A loop controls one agent. A graph coordinates the organisation.

For further implementation lessons, see Five Lessons From Building AI Agents.

What breast-cancer risk prediction teaches every leader

One of the clearest examples of meaningful AI value comes from healthcare, although it is a predictive system rather than an agent. Traditional mammography helps detect signs of cancer that may already be present. Clairity and KUNGFU.AI developed a system using a standard mammogram to estimate a woman’s five-year risk. In May 2025, the United States Food and Drug Administration authorised Allix5, marketed as Clairity Breast, for this purpose.

The breakthrough was not a clever prompt. The model used hundreds of thousands of examinations from 22 clinical sites and was independently tested on 77,511 examinations from 44,112 patients.

The FDA reported a five-year discrimination score of 0.70—not “70% accuracy”—and imposed clear boundaries: the system does not diagnose cancer, replace the radiologist or act as the sole basis for a clinical decision.

The lesson is larger than healthcare. Valuable AI combines distinctive data, rigorous testing, workflow integration and retained human accountability. Trust is what allows value to scale.

Moving from dashboards to always-on intelligence

Traditional business intelligence waits for a person to open a dashboard and interpret what happened. Agents can bring intelligence into the flow of work.

Instead of displaying a backlog, an agent can identify likely causes, prioritise cases, prepare actions and route high-risk exceptions.

Sysco’s enterprise platform combines agents, predictive models, conventional automation and human judgement. It starts with the business problem and chooses the right technology for each step.

This is the shift from systems that report to systems that help the organisation respond.

I use these principles in my own business

In my own work, I use AI workflows to prioritise emails, monitor competitors, identify website improvements and support content research.

Each needs a defined goal, the right information, clear boundaries, output checks and a point at which a person takes over.

That is why I encourage leaders to become “client zero”: use AI on your own work before asking the organisation to change.

The leadership response: apply the 5Ps

The shift towards agents can be understood through my 5Ps AI Readiness Framework.

  • People: Who owns the outcome, challenges the output and remains accountable?
  • Process: Which workflow is being redesigned, rather than simply accelerated?
  • Platforms: Which tools, permissions, monitoring and controls make up the harness?
  • Proprietary Data: Which trusted organisational context gives the agent an advantage?
  • Products & Services: Which measurable customer or commercial outcome will improve?

Organisations also need “questioners”: people who frame the problem, challenge the agent and decide when human judgement matters.

AI should make smaller teams more capable. It should not make accountability smaller.

Read the complete 5Ps framework for an AI-ready business.

Time saved is not value realised

An agent may save a team 1,000 hours, but those hours do not automatically appear in the financial results. Value must be deliberately harvested.

Leaders must decide whether released capacity will reduce backlog, improve service, increase sales, avoid cost or enable a smaller team to handle more work. Measures should move from “hours saved” to cycle time, error rate, cost per transaction, customer satisfaction and revenue.

Invest in one important workflow, capture the benefit, reinvest in better context and controls, then expand.

Start with one workflow, not a fleet of agents

Do not start with an enterprise-wide swarm. Choose one frequent, valuable workflow slowed by information gathering, hand-offs or repetitive judgement.

Map its decisions, data, systems, exceptions and approvals. Build the minimum safe harness, add the right context and define one measurable stopping condition. Test it alongside the existing process.

Only then should you connect more loops into a graph.

Advantage will come from how you combine widely available intelligence with your people, processes, platforms, proprietary data and customer experience.

If you are unsure where your organisation stands, start with the AI Readiness Survey.

Bring this keynote to your leadership audience

The Prompt Is Not the Product is a practical keynote for boards, leadership teams and conferences. It explains what changes when AI moves from answering questions to completing work—and how to create value without losing control.

Audiences leave understanding:

  • The difference between assistants, workflows and agents.
  • The five layers behind dependable agents.
  • Where human accountability must remain.
  • How to select one workflow and begin safely.

Check Mark Kelly’s availability for your event →

The prompt starts the work.

The system around it creates the value.

Sources

Share This Story, Choose Your Platform!