Welcome Back to XcessAI

For the past three years, most of us have interacted with Artificial Intelligence in the same way.

We open ChatGPT, Claude, Gemini or another model, type a prompt, receive an answer, refine the request, and repeat the process until we are satisfied.

Whether we are writing reports, analysing contracts, generating code, summarising research or creating presentations, the workflow has remained largely unchanged. Every step depends on a human deciding what happens next.

This approach has been remarkably successful. It also places humans at the centre of every interaction.

Quietly, that is beginning to change.

Some of the engineers building today's most advanced AI systems have stopped thinking about prompts altogether. Instead, they are designing loops: autonomous workflows that allow AI to plan, execute, verify its own work, and continue iterating until an objective has been achieved.

This represents a fundamentally different way of working with Artificial Intelligence.

A prompt is simply a request.

A loop is a process.

When you write a prompt, the AI performs one task and waits for your next instruction. You remain the project manager, deciding every step of the journey.

A loop works differently.

Instead of telling the AI exactly what to do next, you define the objective, establish the success criteria, and allow the system to determine the sequence of actions required to reach that goal.

Imagine asking an analyst to prepare a board presentation.

Using prompts, you review every draft, request additional analysis, ask for corrections, and approve the final version yourself.

Using a loop, you define what a successful board presentation looks like. The AI drafts it, checks for missing information, improves weak sections, verifies that every required KPI is present, and continues refining the document until the predefined quality threshold has been reached.

Your role changes.

Instead of supervising every step, you design the system that supervises itself.

That is a much bigger shift than simply writing better prompts.

This idea first emerged in software engineering for a very simple reason.

Software is easy to verify. Code either compiles or it doesn't. Tests either pass or fail. Type checks either succeed or generate errors.

That gives AI something extremely valuable: an objective definition of success.

A coding loop might receive a goal such as:

"Fix every authentication test until the full test suite passes with zero lint warnings and no type errors."

The AI writes code, runs the tests, analyses the failures, applies another fix, and repeats the process until either every test passes or a predefined limit is reached.

No human needs to review every intermediate step.

The system knows when it has succeeded.

As AI expands into finance, legal, operations and customer service, organisations are beginning to define similar verification mechanisms for knowledge work.

The same principles apply. Only the tasks change.

Although loops may sound complicated, most of them are built from the same six components.

Understanding these building blocks makes it much easier to design your own.

1. Automation — The Heartbeat

Every loop begins with a trigger.

It might run every morning at 7am, whenever a customer submits a support ticket, when a contract is uploaded, or after a new sales opportunity appears in the CRM.

Instead of waiting for someone to remember the task, the loop knows when to wake up.

Automation is what transforms AI from something you use into something that works for you.

2. Skills — The Playbook

Every organisation has its own way of doing things.

How reports should be written. How emails should sound. How contracts are reviewed. How financial commentary is structured.

Rather than explaining these instructions every time, loops store them as reusable skills.

Think of them as the company's operating manual.

The AI follows the playbook instead of improvising every task from scratch.

3. Sub-agents — The Team

One of the most powerful ideas behind loops is that one AI doesn't have to do everything.

One agent can gather information. Another can perform the analysis. A third can review the output. A fourth can verify compliance or quality.

This mirrors how human organisations already operate.

The person who prepares a report is rarely the same person who audits it.

Separating execution from review significantly improves reliability.

4. Connectors — The Hands

A chatbot can answer questions. A loop can take action.

Connectors allow AI to interact with email, calendars, CRMs, ERP systems, databases, GitHub, Slack, Notion and thousands of other applications.

This is the difference between an assistant that says, "Here is the email you should send," and one that drafts the email, attaches the correct documents, schedules the meeting and updates the CRM automatically.

Connectors turn AI from an advisor into an operator.

5. Verification — The Gatekeeper

This is arguably the most important component of every loop.

Every iteration needs an objective way to determine whether the work has actually been completed successfully.

Did every KPI reconcile? Did the report include every required section? Did the code pass every test? Did the customer receive an answer to every question?

Without verification, AI simply assumes it has done a good job. With verification, it knows whether another iteration is required.

This is what separates genuine progress from endless repetition.

6. Guardrails — The Limits

Finally, every autonomous system needs boundaries.

How many attempts should it make before stopping? How much can it spend? When should it escalate to a human? Which systems is it allowed to access?

Without guardrails, loops can consume thousands of dollars in compute, repeat the same mistake indefinitely, or continue working long after they should have stopped.

The most effective autonomous systems are the ones with the right constraints.

Imagine a CFO preparing the monthly executive briefing.

Instead of manually gathering information every month, a loop could begin on the first business day at 7am.

If inconsistencies remain, the loop continues refining the report. If reconciliation still fails after two attempts, the workflow automatically escalates the issue to Finance.

Nobody needed to write five separate prompts.

The loop managed the process.

This is precisely why many software engineers believe loops represent the next evolution of AI workflows.

It is tempting to think that loops are simply another developer tool. They are not.

Every business process that follows a repeatable sequence is a candidate for a loop.

Anywhere work follows a predictable rhythm, loops have the potential to reduce coordination while improving consistency.

The opportunity is not to replace human judgment. It is to remove the repetitive orchestration that consumes so much of it.

Loops are also capable of amplifying mistakes.

A poorly defined objective will simply produce better iterations of the wrong outcome.

Weak verification allows errors to compound unnoticed. Missing guardrails can turn a useful workflow into an expensive one. And perhaps the greatest risk is assuming that autonomy eliminates responsibility.

It doesn't. The human role simply moves.

Instead of writing every prompt, we define the objectives, establish the rules, design the verification process, and decide when human intervention is required.

In many ways, building loops resembles managing teams. The best managers do not tell people what to do every five minutes. They define goals, establish expectations, monitor progress, and intervene when necessary.

AI is beginning to work the same way.

Prompt engineering taught us how to communicate with AI. Loop engineering is teaching AI how to work without waiting for instructions.

That shift may prove more significant than the next benchmark improvement or the next model release. Because once AI can reliably execute complete workflows, competitive advantage will depend less on writing clever prompts and more on designing intelligent systems.

The organisations that master this transition will build businesses where AI becomes an active participant in day-to-day operations.

The conversation with AI is evolving. The next chapter is teaching it how to work.

Until next time,
Stay adaptive. Stay strategic.
And keep exploring the frontier of AI.

Fabio Lopes
XcessAI

💡Next week: I’m breaking down one of the most misunderstood AI shifts happening right now. Stay tuned. Subscribe above.

Read our previous episodes online!