Artificial intelligence is moving through distinct stages of evolution.
The first big movement was the popularisation of language models and chat interfaces.
For the first time, millions of people could interact directly with an artificial intelligence through natural language. You asked a question, gave an instruction or described a problem, and you got an answer.
Then we moved from chats to agents.
Instead of only answering, models began to use tools, access information, execute tasks, make intermediate decisions and pursue goals with a higher degree of autonomy.
But as agents became more capable, a new challenge appeared.
It was no longer enough to have one agent executing an isolated task.
You had to coordinate different agents, different tools, different sources of information and different levels of autonomy inside the same system of work.
That is the point where the most advanced discussion starts to shift from individual agents to orchestration.
How do you define each agent's role?
How do you coordinate specialised agents?
How do you set quality criteria?
How do you decide what they may decide?
When should they stop?
When do they need to escalate a situation to a person?
How do you evaluate whether their results can be trusted?
And, above all, how do you integrate these agents with the people and the real processes of an organisation?
That is the stage I consider most relevant right now.
Using artificial intelligence is not enough, and neither is building agents on their own.
What matters is building systems in which people and agents work in a coordinated way to produce results.
Following this evolution, studying the most recent work on orchestration and, at the same time, applying these ideas while building my own agent systems, I began to see a very clear connection with my own career.
That was when I realised that many of the competences this new phase of artificial intelligence demands are not entirely new.
The technology is new.
The capability of the models is new.
The speed at which systems can execute, learn and adapt is new.
But the need to understand the work, define processes, set standards, control risk and improve continuously has much older roots.
My career began in quality systems: ISO 9000, process analysis, standardisation and HACCP (Hazard Analysis and Critical Control Points).
In that environment, we learned to look at a process and ask fundamental questions:
Where does the information come from?
What steps are necessary?
What are the critical points?
What can go wrong?
What is the expected result?
How do we know a delivery is correct?
At what moment does someone need to intervene?
What should happen when there is a deviation?
And how does the learning from that deviation get incorporated back into the process?
For a long time, these questions belonged to the language of quality and process management.
Over the years, though, some of these practices started to be associated with bureaucracy.
Talking about procedures, work instructions, controls and standardisation often began to project an image of rigidity and slowness.
In some organisations, the pursuit of agility ended up being confused with the absence of method.
Documentation came to be seen as a burden.
Process came to be treated as a limit on creativity.
And the expectation became that experienced people would simply know what to do.
Now, with the evolution of agentic systems, several of these competences are back at the centre.
We don't need to bring back the bureaucracy of the past, or to produce enormous procedures nobody uses.
These competences matter again because agents can only work reliably when the work is sufficiently understood and structured.
An artificial intelligence agent does not join an organisation already knowing its goals, its criteria, its risks and its way of operating.
It does not know, by nature, which information can be trusted.
It does not know which decisions it may take.
It does not know which results will be considered acceptable.
It does not know in which situations it should keep trying.
It does not know when it needs to change strategy.
It does not know when it should stop and escalate the problem to a person.
And it does not know which lessons need to be incorporated into the system to improve the next deliveries.
All of this needs to be defined.
All of this needs to be tested.
And all of this needs to be improved continuously.
It was working with agents in practice that made this connection even more evident to me.
When I build an agent, I am not just choosing a model or writing an instruction.
I need to determine that agent's function inside the system.
Which sources it may use.
Which tools are available.
What counts as a successful delivery.
What its limits of action are.
Which situations require escalation.
When it should switch context, tool or approach while still pursuing the same goal.
And how its mistakes and successes will be turned into learning.
This is very similar to onboarding a person.
When someone joins an organisation, we need to explain their role, their responsibilities, the sources of information available, the criteria for success and the limits of their decisions.
We also need to teach them to recognise when to ask for help, how to read changes in the environment and how to turn what they learn into improvements for the company.
With agents, several of these needs remain.
We need to define who that agent is within the architecture.
What its purpose is.
Which results it must produce.
Which sources it is authorised to use.
What its guardrails are.
Which decisions it may take.
Which situations it must escalate.
How it should react when the context changes.
And what needs to enter its cycle of knowledge and improvement.
Building trust, however, is different.
When we develop people, we have to consider human, individual elements.
Each person has a history, a motivation, a way of receiving feedback and a speed of learning.
A leader needs to understand how to teach, how to communicate a mistake, how to develop autonomy and how to guide that person along their journey.
With agents, the feedback process can be far more direct and objective.
The error needs to be identified quickly.
The difference between the expected result and the actual result needs to be described precisely.
The instruction needs to be corrected.
The context needs to be adjusted.
A new rule may need to be created.
A test may need to be added.
A source may need to be replaced.
Or the design of the process itself may need to be revised.
An agent does not have the emotional development curve of a person.
But the system built around it has a maturity curve.
And that maturity does not appear just because we use the newest or most powerful model.
It is built through the quality of the orchestration.
The process needs to be designed.
The instructions need to be refined.
The sources need to be evaluated.
The results need to be compared against objective criteria.
The errors need to be classified.
The risks need to be understood.
And autonomy needs to be expanded according to the evidence of reliability.
This dynamic connects directly to the principles of continuous improvement that have run through my whole career.
We define a method.
We execute.
We check the result.
We identify the deviations.
We correct.
We execute again.
When the process shows stability, we standardise.
When we find a new opportunity for improvement, we evolve again.
It is the relationship between PDCA and SDCA.
One cycle to improve the process.
Another to sustain what has already proven to work.
Studying today's approaches to orchestration and applying them in my own projects, I realised this experience lets me see certain challenges very clearly.
Across my career I have worked with quality, processes, business transformation, team development, technology, data, artificial intelligence and the delivery of results.
I learned that an organisational result is never the consequence of a single element.
It is produced by the interaction between people, processes, information, decisions, technology, controls and learning.
Now agents become part of that system.
And that changes profoundly how work needs to be led and executed.
For that reason, I believe professionals and leaders will need to develop a new form of bilingualism.
They will still need to speak the language of people.
The language of motivation, culture, behaviour, communication, leadership, experience and human judgment.
But they will also need to learn to speak the language of agents.
The language of clear goals.
Of verifiable criteria.
Of well-defined contexts.
Of controlled sources.
Of guardrails.
Of escalation conditions.
Of tests.
Of evaluation cycles.
Of memory and structured learning.
And there is a third competence, more important still.
It will not be enough to know how to lead people.
It will not be enough to know how to configure and coordinate agents.
It will be necessary to orchestrate people and agents inside the same system of work.
Because there is no point in building an extraordinary agent if the people in the area don't know how to use it.
There is no point in automating an activity if the rest of the process keeps operating on the old logic.
There is no point in producing an analysis in seconds if nobody knows who validates the result, who takes the decision or who executes the next step.
Orchestration cannot live only inside the technology.
It has to live in the organisation.
Agents need to be connected to people.
People need to understand the agents.
Processes need to define how the two interact.
And decisions need to happen inside a coherent system of responsibilities, criteria and learning.
This is also where one of the biggest traps of this transition appears: the supervision trap.
It happens when we use artificial intelligence to do exactly the same things we already did, only faster.
We ask an agent to execute an activity.
We track each step.
We review the whole delivery.
We redo part of the work.
And we remain responsible for every detail.
In that scenario, artificial intelligence may raise productivity, but it does not truly transform the nature of the work.
The person remains the centre of the execution.
Still supervising everything.
Still living practically the same routine, only managing a larger volume of deliveries.
The real transformation happens when we redesign the work.
When we define what will continue to be done by people.
What will be delegated to agents.
Which results need to be verified.
Which processes can operate with autonomy.
Which risks demand human control.
And under which conditions a person needs to take the activity back.
Trust does not mean accepting whatever an agent produces.
Trust means knowing the system well enough to know when it can work autonomously and when it needs intervention.
It is that structured trust that allows the routine to change.
Instead of using agents merely to do more tasks, we begin to stop executing certain activities.
We stop controlling steps that have already shown stability.
We stop being the operators of the whole process.
And we start to act as architects, evaluators and orchestrators of the system.
This evolution, from language models to agents and from agents to orchestration, is what led me to start writing here.
I am studying this transformation, and I am also applying these principles: building agents, evaluating their results and learning, in practice, how to integrate them with processes and people.
These essays will be dedicated to two big questions.
How to lead in the world of artificial intelligence.
And how to work in the world of artificial intelligence.
They are not the same thing.
Leading means understanding how to build teams made of people and agents.
How to define responsibilities.
How to establish governance.
How to develop trust.
How to evaluate results.
How to coordinate different capabilities.
And how to transform an organisation.
Working in this new world means understanding how our own routine changes.
How we use agents without endlessly increasing our supervision load.
How we build systems that work with us.
How we stop executing certain activities.
How we expand our capacity.
And how we produce results through the combination of human judgment and agentic execution.
That is the transformation we are going to talk about.
Tools, models and prompts are part of it.
The larger subject is how to build a new way of working.
How to turn agents into reliable capabilities.
How to onboard those capabilities.
How to evaluate their results.
How to correct their mistakes.
How to standardise what works.
How to improve continuously.
And how to integrate people and agents inside one system of work.
We are entering a phase in which a competence that seemed, for a while, to be losing importance will become decisive again.
The ability to see the work.
To understand the process.
To set criteria.
To recognise critical points.
To control risk.
To build trust.
And to turn learning into a new standard.
This is not the bureaucratic standardisation of the past.
It is a living standardisation, adaptive and continuously refined.
A structure that allows people and agents to work together with safety, quality, efficiency and autonomy.
It was this connection between the beginning of my career and what I am building today that made me realise I needed to talk about this.
Because it connects quality, processes, continuous improvement, technology, artificial intelligence, management and leadership.
And because the next frontier will not be defined only by the capability of the models.
It will be defined by our capacity to organise the work those models do.
The future will not be only human.
It will not be only agentic.
It will be human and agentic.
And someone will have to conduct that orchestra.