From talking to AI to directing teams of AI agents

Imagine two scenes.

In the first, someone opens an AI application on their phone and asks:

“Can you summarise this contract?”

The AI responds. The person reads the answer, decides what to do with it and manually takes the next step.

In the second scene, the person sets an objective:

“Analyse these contracts, compare them with our standard agreement, identify the clauses that depart from company policy, verify the legal references and prepare a report for my review.”

The system opens the authorised documents, classifies their contents, consults the available sources, runs comparisons, distributes parts of the work, reviews the results and produces a report highlighting the issues that require a human decision.

Both scenes involve artificial intelligence. But they represent radically different ways of working.

The first produces an answer.

The second attempts to complete a task.

The phone is a window, not the frontier

We tend to identify artificial intelligence with the place where we talk to it: an application on a phone, a website or a text box.

But the screen is only the interface.

A phone can be used to ask a simple question. It can also serve as the control panel from which work is launched, supervised or approved while it runs on another computer, on a server or inside a protected environment.

The important distinction, therefore, is not whether we use AI from a phone or a computer. Nor does it depend solely on which model is the most powerful.

The distinction is whether the system can only talk or whether it can also act.

What is an AI agent?

An agent is not necessarily a new kind of artificial intelligence. It is a system built around a model to turn its outputs into a sequence of actions.

Put simply:

A conversational AI responds to a request.

An agent receives an objective, decides some of the steps, uses tools, observes the results and adapts what it does next.

The objective does not originate in the AI itself. It is set by a person, an organisation or a program. What changes is that the system has some latitude to decide how it will try to achieve that objective.

Agentic AI is not an all-or-nothing category either. There is a continuum between fully programmed workflows and systems with greater freedom to select their next action. The OECD describes agency as a spectrum, rather than a switch that is simply on or off.

In a traditional workflow, the route is written in advance:

1. Receive a document.
2. Extract its data.
3. Apply a set of rules.
4. Generate a report.

In a more agentic system, the model may decide that it needs to consult another source, run an additional check, seek help from a specialist agent or revisit a conclusion that does not fit the evidence it has found.

The most reliable systems tend to combine both approaches: flexibility where it is useful and deterministic rules where the process must not be left to improvisation.

The harness: what turns a model into a working system

The model is only one part of the whole.

To act, it needs what the industry commonly calls a harness: an execution layer that connects the model to the rest of the system.

That harness may be responsible for:

• preparing the information the model receives;
• making authorised tools available to it;
• executing the operations it requests;
• preserving the state of the task;
• returning the results;
• detecting errors and allowing further attempts;
• applying limits on time, cost and permissions;
• requesting human authorisation;
• recording what happened;
• stopping the process when a limit is reached.

We can think of the model as an engine. The harness is the complete vehicle: steering, brakes, instruments, seat belts and the rules governing where it may travel.

A more powerful engine does not, by itself, guarantee a better vehicle.

Tools are the hands

An isolated model can generate text, images or code. But to work on the digital world, it needs tools.

A tool may enable it to:

• search approved sources;
• read documents;
• query a database;
• create or modify files;
• run code;
• use a business application;
• update an internal system;
• assign an operation to another agent;
• request a person’s approval.

The model does not normally carry out these actions itself. It requests an operation through a structured instruction, and the system checks whether that operation is permitted before executing it.

This distinction matters: an agent’s real capability depends not only on what the model “knows”, but also on the tools, data and permissions placed at its disposal.

From explaining code to working on a project

Software development makes this change particularly easy to see.

A conversational AI can explain how to fix an error and provide a code fragment.

A system such as Codex can inspect a project’s files, locate the source of the problem, modify the code, run tests, observe a failure, correct it and present the changes for review.

The same logic can extend to many other fields:

• not merely explaining a spreadsheet, but analysing it and preparing a new version;
• not merely summarising a regulation, but identifying which company processes it may affect;
• not merely describing what a dashboard should look like, but programming a tailored prototype;
• not merely suggesting a search, but carrying it out, comparing the sources and preparing a documented report.

The leap is from generating content to operating within an environment.

AI can also create its own working tools

One of the most transformative possibilities is that these systems need not be limited to using existing software.

A coding agent can build a small application for a specific need: a clause comparison tool, an information extractor, a specialist calculator, a monitoring dashboard or a connector between two internal systems.

Once reviewed and validated, that tool can be incorporated into the workflow and used by other agents.

This lowers the cost of developing software tailored to very specific needs. Processes that would never have justified months of development may begin to be addressed through small applications built around how an organisation actually works.

The business advantage, therefore, may not lie solely in buying access to the best model. It may lie in building the best system around it: appropriate data, proprietary tools, clear processes, useful memory and effective controls.

Working memory and long-term memory

Completing a complex task requires more than answering each individual question correctly. The system must preserve the thread of the work.

Working memory includes what the agent needs during the current task:

• the objective;
• the plan;
• the relevant documents;
• the operations already carried out;
• the results obtained;
• the errors encountered;
• the decisions still pending.

It is like the desk on which the materials for a matter are laid out while the work is under way.

Long-term memory is different. It can preserve information between sessions: earlier decisions, preferences, projects, relationships between documents, internal procedures or relevant experience.

But it is neither human memory nor a continuous consciousness. It generally consists of files, databases, records and summaries that the system retrieves when it considers them useful.

Nor does remembering more always mean working better. Memory that is incorrect, outdated or associated with the wrong matter may be more dangerous than no memory at all.

Professional memory therefore requires provenance, dates, confidence levels, access rules and mechanisms for correcting, replacing or deleting information.

Can several agents work together?

Yes. It is already possible to use a coordinating agent that divides an assignment and delegates parts of the work to other agents.

The official Codex documentation, for example, describes the use of specialised subagents that work in parallel and return their results to a coordinator.

We might configure:

• one agent to find information;
• another to analyse documents;
• another to verify figures;
• another to check sources;
• another to look for contradictions;
• a coordinator to integrate the results.

It is important, however, to understand what “team” really means in this context.

We are not talking about digital people with independent professional judgement. Often, these are separate instances of the same model, each operating with different instructions, tools and context.

Nor is it true that adding more agents always produces better results. Available research suggests that parallel work can help when tasks can be separated, but can also increase conflicts, costs and the propagation of errors when a problem requires a tightly coordinated sequence. Google Research has documented these differences.

A team of agents adds execution capacity. It does not automatically add responsibility, sound judgement or a shared understanding of the objective.

What could this mean for a business?

Consider a company assessing whether to enter a new market.

An agentic system could:

1. Gather economic and sector information.
2. Identify the relevant regulation.
3. Compare competitors and suppliers.
4. Review contracts and risks.
5. Create a small program to organise the data.
6. Check its conclusions against approved sources.
7. Prepare different scenarios.
8. Present the outstanding decisions to management.

It could also help maintain an up-to-date map of regulatory obligations, prepare draft reports, detect irregularities in internal processes, turn meetings into tasks or create dashboards tailored to each project.

None of this necessarily requires artificial general intelligence. It requires combining today’s models with programs, tools, data, memory and an appropriate architecture.

The key is to identify which parts of the work may be automated, which require verification and which must remain human decisions.

And what could it mean for a law firm?

The change may be even more profound in a law firm, because so much legal work takes place through documents, information, rules and procedures.

When a new matter arrives, a system could:

• inventory and classify the authorised documentation;
• extract dates, people, amounts and events;
• build a chronology;
• locate official legal sources;
• compare documents with templates and precedents;
• flag contradictions or missing information;
• prepare questions for the client;
• produce a first draft of the analysis;
• verify that citations lead to the correct sources;
• present the lawyer with the issues requiring professional judgement.

The lawyer would still decide the strategy, assess the context, advise the client and bear responsibility for the outcome.

That boundary is essential. European and national professional bodies emphasise that using AI does not displace the lawyer’s responsibility. See, for example, the CCBE guide for lawyers and the guidance from the Norwegian Bar Association.

The principle can be expressed simply:

Autonomy to prepare; human authorisation to commit.

An agent may research, organise, compare and propose. But whenever an action affects a client, a third party, a deadline, a right or an obligation, there must be an identifiable human decision.

What is already possible—and what is not

Systems already exist that can:

• use tools;
• run code in controlled environments;
• work on repositories and documents;
• operate certain interfaces;
• continue tasks for extended periods;
• retrieve persistent information;
• delegate subtasks;
• coordinate work in parallel.

This makes it possible to build highly capable digital teams for specific kinds of work.

But it is not the same as having an autonomous company inside a computer.

Today’s systems still struggle with ambiguous objectives, incomplete information, conflicting instructions, very long tasks and situations in which there is no clear way to verify whether the result is correct.

They work best when the task is digital, well defined, based on known sources, supported by verification criteria and designed so that errors can be reversed.

Agentic AI is not necessarily AI with a better understanding of the world. It is, above all, AI with a greater capacity to act upon it. A reasoning error can therefore cease to be merely a wrong answer and become a sequence of wrong operations.

Where the technology is heading

The development appears to be moving in several directions at once:

• from conversations to complete tasks;
• from general assistants to specialised agents;
• from a single session to systems with continuity;
• from one isolated agent to coordinated teams;
• from using existing programs to creating tailored tools;
• from waiting for an answer to supervising processes that work in the background;
• from controlling every click to defining objectives, limits and approval points.

The human role changes as well.

The person stops performing every operation manually and instead designs the work, defines the limits, supplies context, supervises exceptions and takes responsibility for decisions that have consequences.

This does not remove the need for professionals. It changes where they add value.

The real question

For the past few years, we have asked which model gives the best answers.

The next question will be broader:

Which system works best?

That will depend on the model, but also on the quality of its tools, the memory it retains, the sources it uses, how tasks are distributed, who can grant permissions and when it must stop to request a human decision.

Giving an answer is generation.

Completing a task is operation.

Creating a tool is expanding capacity.

Coordinating agents is organising work.

That is where agentic AI truly begins.

And that transition may profoundly change businesses and law firms—not because machines become people, but because we are beginning to build systems capable of turning a human instruction into organised work.

The question is no longer only what an artificial intelligence can say.

The question is what we allow it to do, with which tools, under which rules and under whose responsibility.