Module 8Academy

Building agents for operations

An agent that completes a multi-step task, with limits, checks and a person at the end.

Building agents for operations is a training day in which the team designs an agent that handles a task in several steps. The agent gets a limited set of tools, clear boundaries and a checkpoint for every action with consequences. The team builds one working agent on a task of its own and learns when an agent is the right choice, and when a fixed chain of steps is better.

After this module the team can
  • 01The team describes an agent as a loop of observation, decision, action and check, and can fill it in for its own task
  • 02The team decides the agent’s tools and writes down which actions need human approval
  • 03The team builds one working agent on a bounded task and tests it on real exceptions
  • 04The team can explain when an agent is the wrong choice and a fixed chain fits better
Agenda
  1. 01

    What an agent is, and what it is not

    An agent chooses its next step within limits. A chain follows a fixed path. With the participants we decide which of the two fits their task.

  2. 02

    Tools and boundaries

    Which systems the agent may use, with which rights, and which actions only with approval. We write the boundaries down before we build.

  3. 03

    Building and testing

    We build one agent on a task of the team’s own. Then we test on the exceptions the team knows, and note what the agent gets wrong.

  4. 04

    Watching

    Logging of every step, a measure of the outcome and an agreement on who switches the agent off when it drifts.

What makes an agent different

An automation carries out predefined steps. An agent is given a goal and a set of tools, and chooses itself which step to take next. That lets it finish a task that cannot be caught in a chain beforehand, for example a client question that touches three systems and an exception nobody foresaw. That is the promise, and at the same time the risk.

The question for a team is therefore not whether an agent is clever enough, but whether the task needs freedom of steps. Many tasks sold as agents are in reality a chain with one model inside. Those are more reliable, cheaper and easier to check.

Tools and boundaries

An agent is only as safe as the tools it has. Each tool is an action: reading a file, drafting an email, changing a record, making a payment. The training starts with the list of tools and removes everything the task does not need. Then comes the question per tool: may the agent do it on its own, or only after a person approves?

The model does not remember those boundaries by itself. They must be in the instruction and in the rights of the connection. An agent allowed to make a payment will do so on a day when the instruction is unclear. That is why actions with consequences outside the system always stay behind an approval.

One task, not a platform

The training builds one agent on one task of the participants. That is deliberate. An agent for the whole company is a months-long project and teaches the team little about its own limits. An agent for one task, such as reviewing incoming supplier invoices or preparing client replies, shows where the model is good and where it will slip.

Which platform or framework we use depends on the company’s environment and its management requirements. That choice is not the learning goal; the learning goal is the design.

Testing on the exceptions

An agent is not tested on the average day but on the difficult one. With the team we collect the ten exceptions that occur most often or cost the most, and run the agent through those first. What it gets wrong is the information the team needs. An agent that handles only the easy cases well is not ready for production.

Logging is not optional. Every step, every tool and every decision is recorded, so the team can later see why the agent did something. Without that log every error is a riddle.

When a chain is better

The end result of the training is not always an agent. If the task consists of five fixed steps and the exceptions are known, a chain with one model step is more reliable. The team learns that distinction explicitly and uses it as a test for every next agent wish. That is one of the most valuable outcomes of the day.

Frequently asked questions

  • Is an agent the same as a chatbot?

    No. A chatbot answers. An agent acts: it queries systems, takes steps and can do something. That is why an agent needs more boundaries and checks.

  • Can we build a production-ready agent in one day?

    A working version on a bounded task, yes, which must then still be tested and watched. A production-ready agent is only ready after that test, and the test takes longer than a day.

  • Does this fall under the AI Act?

    That depends on what the agent does and for whom. An agent that decides about people can fall under the high-risk rules. Module five and the glossary give the test.

Which team needs this module?

For up to 12 participants €3,950 excl. VAT. In the conversation we say whether the module fits, or whether a build day or the adoption programme is better.