Dev48
Language
  • About
  • Services
  • Industries
  • Technologies
  • Articles
  • Contacts
Book a call
    Home/Articles/Dont make the agent do everything
Dev48
info@dev48.org+375 29 320-40-96WhatsApp@isky_by

© 2026 · All rights reserved.

Don’t make the agent do everything

Source: Atos•November 17, 2026 at 12:06 PM

Don’t make the agent do everything

Giving an agent a central role doesn't mean asking AI to do every step in a process. Peter Kalmijn argues that the right balance lets agents interpret and coordinate work while established software continues handling tasks that require consistent execution. 

Giving an agent a central role doesn't mean asking AI to do every step in a process. Peter Kalmijn argues that the right balance lets agents interpret and coordinate work while established software continues handling tasks that require consistent execution.

Imagine we are designing an AI agent to handle a complicated customer request. It reads the email, examines the attachments and works out what information matters. It pulls information from several internal systems and decides what to do next.

Then it calculates the price – again.

Why? We already have a pricing service.

This situation is the kind of moment I find particularly interesting when thinking about agentic architecture. Once an agent demonstrates that it can reason and use tools, it becomes surprisingly appealing to hand it work simply because we can. Eventually we may find ourselves asking an AI system to think about something the organization already knows how to do. Over and over again.

A process contains different kinds of work

Let's stick with our customer request. Understanding what the customer wants may require reasoning and interpretation. The customer's language will not necessarily match the internal terminology. Information may be split across the message and several documents, and something important may be missing.

AI can do useful work here.

Once the agent understands the request, it may need a price. The organization already has a service that takes defined inputs and returns a calculation based on approved logic. At this point, the architecture has a clear choice: calculate using AI or use the approved deterministic logic.

We can provide the pricing logic to the model and let it reason through the calculation, or the agent can collect the required inputs and call the pricing service. I prefer the second approach unless there is a good reason to do otherwise.

The agent remains central to the process. It interprets, coordinates and determines which capability it needs, while the established service remains responsible for the calculation it was built and tested to perform. Each component has a clear job, which also makes it easier to determine what happened when something goes wrong.

Old software does not stop working because AI arrived

Proven systems matter, especially in large organizations, because they rarely start from a blank sheet of paper. There are applications, databases, APIs, workflow systems, rules-based decision engines and custom-built services. Some are modern. Some are proven technology. Some definitely aren't, and probably should have disappeared years ago.

Age and usefulness are separate questions, though. A service that reliably performs a calculation or executes a business rule may remain perfectly valuable inside an agentic process. The agent only needs to know when and how to use that service as a tool.

That opens an interesting architectural possibility. Agents can provide flexibility around capabilities that themselves should remain deterministic. Users and customers can use natural language and unstructured information while established systems keep doing the work they were designed for.

This possibility comes with a warning. Putting an AI agent in front of a poorly understood legacy application does not suddenly make the underlying application good. Security, interfaces, data quality, ownership and maintainability still matter.

AI is no excuse for bad architecture.

Look for the handover

Where should the agent stop reasoning and use another component? There is no universal line, so I look at the nature of the problem.

At the beginning of a process, we may be asking:

  • What is happening here?
  • What information matters?
  • What should I do next?

Later, once this ambiguity is resolved, the question may become:

  • Given these validated facts, what do the approved rules say?

That change in the question's nature is a useful candidate for a handover.

Consider an eligibility process. AI might examine supporting documents, extract relevant information, and highlight eventual gaps. Once those facts have been checked, an existing decision service can deterministically apply the eligibility rules. We can then see what happened at each stage: what information the AI identified, which facts entered the service, which rules were applied and which result came back.

This separation also helps us know where to start when a result is wrong. Did the AI extract the wrong information? Did it choose the wrong tool? Were incorrect inputs passed to the service? Is the deterministic business rule itself wrong?

If the model reasons through the whole process in one opaque chain, those distinctions can become much harder to see.

AI can help with the rules too

Another role for AI particularly interests me. Rules often originate in legislation, policies, contractual material or lengthy technical documents. Understanding that material can require substantial interpretation to locate relevant passages, compare versions and determine how different decisioning requirements relate.

AI can assist with that very knowledge-intensive work. Once the resulting logic has been validated, approved by the analyst, linked to the source and implemented, executing it is the next step. A decision service can apply the rule repeatedly without reconstructing the reasoning that produced it.

This is an important distinction I expect us to encounter more often as agents become part of enterprise architecture. AI-assisted reasoning can help us design a deterministic decision service, so AI doesn't have to reason again every time we need the decision outcome.

Saying no in the right places

An AI agent with well-defined boundaries can still have substantial autonomy. It can decide what information to retrieve, adapt as the situation changes and choose the right tools to reach its goal. At specific points, the architecture can require it to call an approved service or involve a person.

Those boundaries give us more confidence in where autonomy is useful because we have also considered where it isn't.

Maximizing the percentage of a process handled by an AI model is not an end in itself. If the production architecture has less AI than the prototype, it costs less to operate, behaves more consistently and is easier to investigate. I would certainly prefer that architecture.

So when designing an agentic system, I would keep returning to one question:

Where is agentic reasoning actually needed?

The answer may lead us to use AI extensively. And it may also lead us to say no to AI.

Good architecture needs both answers.

"We’re looking forward to our sponsorship at Adopt AI 2026, December 3-4, in Paris. Each Thursday leading up to the event, Agile & Digital Enterprise Architect Peter Kalmijn, a member of the Future Makers Research Community (FMRC) at Atos", will explain why good AI architecture starts by saying

"We’re looking forward to our sponsorship at Adopt AI 2026, December 3-4, in Paris. Each Thursday leading up to the event, Agile & Digital Enterprise Architect Peter Kalmijn, a member of the Future Makers Research Community (FMRC) at Atos", will explain why good AI architecture starts by saying

← All articles

What this article says

More in Cloud & Infrastructure

All →
5 ways Amazon is supporting US communities in 2026, from free meals to affordable homes
Amazon

5 ways Amazon is supporting US communities in 2026, from free meals to affordable homes

How low-SWaP radar is making sophisticated sensing smarter
IBM

How low-SWaP radar is making sophisticated sensing smarter

Amazon to invest $7 billion in renewable energy in India by 2030
Amazon

Amazon to invest $7 billion in renewable energy in India by 2030

Amazon sets goal to cut electricity carbon intensity 65% by 2030
Amazon

Amazon sets goal to cut electricity carbon intensity 65% by 2030

Spacelift vs GitHub Actions: Key Differences
Spacelift

Spacelift vs GitHub Actions: Key Differences

Inside Amazon’s new Tokyo office with views of Japan's iconic Tokyo Tower
Amazon

Inside Amazon’s new Tokyo office with views of Japan's iconic Tokyo Tower

More from Atos

You can afford the AI pilot. Can you afford the production system?
Atos

You can afford the AI pilot. Can you afford the production system?

Threat actor playbooks: Who is targeting the IT supply chain & how
Atos

Threat actor playbooks: Who is targeting the IT supply chain & how

Just because AI can do it, does that mean it should?
Atos

Just because AI can do it, does that mean it should?

Cyber at machine speed: Can your security program keep up?
Atos

Cyber at machine speed: Can your security program keep up?

Move from prediction to resilience: Navigating the new cyber equilibrium
Atos

Move from prediction to resilience: Navigating the new cyber equilibrium

Securing AI workloads across the hybrid and multi-cloud AI supply chain
Atos

Securing AI workloads across the hybrid and multi-cloud AI supply chain