One agent pulls. Another writes.

One agent pulls. Another writes.

Next week Paul is teaching agents to students at a business school. The demo is a gym with nine hundred members and a lapsing-member problem. He has been building it from his phone, and the bit worth stealing is not the gym. It is the shape of the team.

Three agents. Carlos pulls the data and computes the numbers. Enrique reads what Carlos found, decides what to do, and writes the notes. Cato doubts both. Each one has a rule about what it cannot touch. Carlos cannot write prose. Enrique cannot touch the door log. Cato does nothing but ask questions.

The obvious way to build this is one agent that does everything. Pull the numbers, decide who lapsed, write the email. That is how most people start.

Here is the problem with an agent that does everything. When it writes a number in a sentence, it is now doing two jobs at once. It is computing and describing. If the description is wrong, you cannot tell whether the maths was wrong or the phrasing was wrong. If you fix the phrasing and the maths was also wrong, the number looks right and is not.

So Carlos never describes a number. He runs the code, copies the result as a figure, and hands it over. Enrique takes what Carlos gave him and writes around it, never through it. When the note says a member has not scanned in for forty-two days, that forty-two came straight from Carlos. Enrique did not round it or phrase it or decide what it meant. He just used it.

We run the same split on our own research desk. A script runs the tools and pastes the results into each research card exactly as they came out. A gate then refuses the card if it carries a big number that is not in the pasted block, so the agent writes around the numbers and never through them. A different agent then checks every card.

The students will also build the loop that closes. A gym email agent that stops at sending the email has not finished its job. It finished the task, but the task was not the job. The job was to get members back to the gym. So Paul planted a file of pretend returns and the agent reads it two weeks later and reports the percentage who came back. That readout is what the students see an agent's job actually is.

Then there is Cato, the fresh-context checker. It reads what the other two produced, knows nothing about how they got there, and asks whether the output makes sense. If Carlos says forty-two days and Enrique says a month and a half, Cato will notice. Nobody plants mistakes for Cato to find, and nothing in the week is built to fail on purpose. Paul's line: it is hard enough to get things to work properly.

Four questions decide whether an agent is ready. What does it produce. Where does that thing land. How does it know it is done. What does the handoff look like. When students read the spec that Enrique runs on, they are reading the same document that asks those four questions and answers them.

An agent is easy to build. You can build one from your phone. Building agents that consistently do what you want them to do is a skill and a practice. As Paul put it, an agent that pulls information and cannot write comes from practice, curiosity and discipline. And the first lesson of that practice is to stop one agent doing two jobs at once.

Lena