Free Agent or Fixed Workflow: When an Agent Should Decide for Itself

Monday morning at a plant builder, an invented but familiar example. The head of sales no longer wants to put together the weekly report every Friday herself. At the same time, a project manager asks what was agreed at the acceptance test of the filling line P-2291 and which of it was not met. A colleague gives both of them the same advice: an agent can do that now.

That sounds like an answer, but it is only the start of a question. An agent can work in two very different ways. It can be handed a task and decide for itself how to approach it: what to look up first, which tool to pick, when it is done. Or it can be one part of a fixed sequence in which somebody decided beforehand which step comes when, who has to approve, and what happens at the end. In the agent list of TheroAI, every entry carries one of two labels: “Frei” or “Ablauf”.

This article is a decision aid for exactly that choice. Our thesis is plain and easy to apply: freedom is a resource, and it has a cost. A free agent is flexible, but a little unpredictable every single time. A fixed workflow is dependable, but it can only do what somebody wrote down in advance. So the question is not how clever the task is. It is how much freedom the task can afford.

Two Ways to Hand Over Work

Let us start with terms, so it is clear what we are talking about. By a free agent we mean an agent that you give a task and some tools, and that decides on its own, during the conversation, in which order to use them. In the agent list you recognize it by the label “Frei” and by the trigger “Wenn jemand mit dem Agenten chattet” (when somebody chats with the agent). A workflow is a fixed series of steps: a trigger, then steps in which an agent summarizes or drafts something, an action looks something up or sets something off, and, where needed, an approval step at which a person decides. A workflow can be started by hand, by a submitted form or on a schedule, for example every weekday at 9:00.

Enlarge

The most important difference is who decides the order. With a free agent, the agent decides while it works. With a workflow, you decide before anything runs. The graphic below puts the two side by side.

Enlarge

Almost everything else follows from that one difference, as the comparison shows.

One note on naming: the German interface says “Frei” and “Ablauf”. In this article we speak of free agents and of workflows. They mean the same thing.

This Is Not a Question of Intelligence

People facing this choice for the first time often ask: is the task hard enough for a free agent, or will a workflow do? That is the wrong yardstick. A hard task that looks the same every week belongs in a workflow, because the difficulty then sits inside one step and not in the order of steps. An easy task that looks different every time cannot sensibly be poured into fixed steps at all.

Four questions help more, and none of them is about the intelligence of the model. They are about the task and its surroundings:

• Repeatability: How often does the task show up in the same form?
• Risk: What does a mistake cost, and can it be undone?
• Checkability: How quickly can a person tell that the result is right?
• Variability: How much do inputs, scope and goal change from case to case?

Each of these questions pulls in one direction. The graphic sums up what speaks for which side.

Enlarge

The Four Criteria in Detail

Repeatability

The more often a task appears in the same form, the more it pays to write the route down once. The weekly sales report is the textbook case: same sources, same structure, same recipients, every week. Briefing a free agent from scratch each time would waste attention. A workflow turns compiling the report into a habit that nobody has to maintain.

The reverse holds too. A question you ask today for the first and perhaps the last time does not justify a workflow. Whoever builds one invests time in something that will never run again.

Risk

This is about the cost of a mistake, and above all about whether a mistake can be taken back. A wrong summary for your own use costs you, at worst, a second attempt. A wrong email to a customer is out the moment it is sent. We described in a separate article why sending at TheroAI only happens after approval.

For the choice between free and workflow, that means: the higher the risk, the more important fixed checkpoints at fixed places become. A workflow can contain an approval step that nobody can forget or bypass, because it is part of the structure. In TheroAI you can name the people who approve and set a deadline, 72 hours in our example, and a switch for the four-eyes principle makes sure that the person who started the workflow and the person who created it cannot decide themselves.

That does not mean a free agent has no checkpoints. Here, too, you can set for each tool whether it is allowed, allowed only after asking, or blocked, and writing tools are marked accordingly. The difference is reliability. In a workflow, the checkpoint sits at a specific place in the structure. With a free agent, it depends on you having remembered to choose the right setting for every tool. More on this under tools and permissions.

Checkability

This question is the one most often overlooked: how quickly can you tell from the result whether it is right? An answer with a table and numbered source references can be checked in minutes, because for every figure you can look up where it comes from. A text that contains a legal assessment cannot be checked in seconds, however convincing it sounds.

The rule that follows: high checkability allows freedom. If you can judge the result quickly, the agent may choose its own route, because you control the end and not the path. Low checkability demands structure. If you cannot tell by looking whether the result is right, you at least need the intermediate steps: which source was read, which step decided what. A workflow provides that by its nature, and the audit log in TheroAI records time, process, action, status, duration and person for every run and can be exported as a CSV file.

A practical note: as soon as logs make statements about individual employees possible, for example who started which run and when, the co-determination rights of a works council may be affected, in Germany for instance under section 87 (1) no. 6 of the Works Constitution Act. This is not legal advice. Clarify the individual case with your legal counsel and your works council, ideally before the start and not afterwards.

Variability of the Task

The fourth criterion is often confused with repeatability, but it is a different one. A task can come up often and still look different every time: customer inquiries arrive daily, but each one is about something else. Conversely, a task can be rare and still always have the same structure.

When variability is high, a fixed workflow fails, and not for technical reasons but because of the writing down: you cannot define the steps because you do not know in advance which ones will be needed. This is exactly where a free agent has the advantage. It looks things up, reads, discards, searches differently and adapts to the question.

Six Tasks from Everyday Office Work

Let us apply the four criteria to six tasks. The ratings are our own estimate for a typical company and not a measurement. For you they may come out differently, and that is the point of the exercise. The names and numbers come from the invented demo documents on our industry pages, for example for plant engineering, and are explicitly examples, not customer cases.

If you put variability and risk on a pair of axes, four fields appear.

Enlarge

Bottom left: fixed workflow. The weekly report and the complaint are recurring, uniform and low in risk, because nothing leaves the company. Here the workflow is the normal case. The report can start on a schedule, the complaint through a form. Nobody has to explain anything, nobody has to kick anything off.

Top left: workflow with approval. The reply to a customer about error E-217 on filler FL-200 looks structurally the same every time: read the inquiry, look up the error description and earlier cases, draft the reply. But it goes outside. So an approval step belongs right before sending, showing the full draft with the sources named. Checkability is high, because the person reads a concrete text and sees where it comes from.

Bottom right: free agent. The question of what was agreed at the acceptance of P-2291 and which of it was not met is different every time. The agent has to search, read and compare. Nothing goes outside, and the answer comes with numbered sources, so you can verify every statement against the acceptance record. A fixed workflow would be not only unnecessary here but a drag: every new question would need new steps. The same goes for the supplier comparison, where the agent pulls in different documents depending on the supplier.

Top right: the hard case. For the change to the framework contract with TP-118, everything comes together: the task is variable, the risk is high, and checking it takes expertise. Nobody can write a workflow for it, because every contract is different. Letting a free agent work blindly would be reckless. The defensible solution is the least spectacular one: the agent reads, compares and summarizes, but it gets no writing tools. The decision stays with the person who has the expertise. That is not a weakness of the system but a division of labor.

Hybrids: Freedom in the Right Place

Most tasks in an office do not sit neatly in one field. That is why the most important insight of this article may be this one: free and workflow are not mutually exclusive. A workflow may contain steps in which an agent works with some freedom. The structure dictates what happens in which order. Inside an agent step, the model decides how to word things and what to weigh.

One example from the product is the workflow “Support-Anfrage beantworten” (answer a support request). It starts when somebody fills in a form. The first step is an agent that summarizes the request. Next, an action searches the knowledge base. A second agent drafts the reply. Then comes an approval step, and only after that does the email go out.

Enlarge

The graphic shows where the freedom sits. Two of the six stations are open, namely the two agent steps. Everything with consequences is fixed: the order, access to the knowledge base, the approval and the send. The agent may word things cleverly. It may not decide whether anything gets sent.

Enlarge

When does freedom inside a workflow make sense? Five places come up again and again:

• When summarizing free text, such as a customer inquiry or meeting notes, because free text never looks the same twice.
• When drafting a text that a person will read afterwards.
• When sorting into categories, if the categories are given and the agent only assigns.
• When picking relevant passages from a larger body of material.
• When phrasing follow-up questions because details are missing in the form.

The counterparts that should stay fixed in the workflow are just as clear: everything that goes outside, everything that changes data, and everything that concerns money or commitments.

A second hybrid runs the other way: a free agent whose tools are deliberately limited. It may read, search and research, but everything that writes is set to “ask first” or “blocked”. That creates freedom in the conversation without anything irreversible happening. Tasks such as the supplier comparison or a contract review are typical candidates.

When a Workflow Pays Off: A Worked Example

A workflow costs build time. When does that pay off? The following calculation is explicitly an example with round assumptions and not a measurement from a customer project. It is meant to show how to think about it, not which numbers apply to you.

Why review time is higher with a free agent is quickly explained: because the route differs each time, you have more to retrace. In a workflow you know the shape and only check the places where there was freedom. Both values are assumptions that you should replace with your own observations.

From this follows: after n runs, the free agent costs 6 times n minutes of review effort. The workflow costs 90 plus 2 times n minutes. The two curves cross at n equals 22.5. From the 23rd run on, the workflow is ahead, which at 20 runs a month means during the second month. After four months and 80 runs, it is 480 minutes for the free agent against 250 minutes for the workflow. The difference is 230 minutes, almost four hours.

Enlarge

The calculation shows two things. First, a workflow is not an investment that pays off immediately. With few runs, the free agent is the cheaper choice, and that is perfectly fine. Second, saved time is not the only reason for a workflow. Where risk is high, we would build one even if the curves never crossed, because a fixed checkpoint is worth something that does not show up in minutes.

From Free Agent to Workflow: A Path, Not an Either-Or

The good news is that you do not have to make the decision once and for all. Many workflows grow out of free agents that have proven themselves. One path that suggests itself:

1. 1.Try it freely. A new task, an unclear route: give it to a free agent, with reading tools and source references. Check every result yourself.
2. 2.Spot the pattern. When you notice around the fifth or sixth time that the agent always takes the same steps in the same order, that is your cue. The number is a rule of thumb, not a rule.
3. 3.Build the workflow. Transfer the steps into a workflow. Leave agent steps open where free text occurs, and fix everything else. Try it in the preview before it runs for real.
4. 4.Set checkpoints. Work out which steps act on the outside world and put an approval in front of them. Decide who approves and how long the deadline is.
5. 5.Observe and adjust. Look at the log. If certain mistakes keep happening at the same place, change that step. If the task changes fundamentally, go back to step 1.

The last point matters. A workflow is not a monument. When the task changes, it is perfectly legitimate to open it again or to return to the free agent.

Typical Wrong Turns

The thoughts above point to a few patterns in which the choice often goes wrong.

The third row deserves attention. An approval is only as good as the attention of the person who gives it. Anyone who clicks away thirty cards a day does not have a safety net but a ritual. The consequence: use approvals sparingly and exactly where something can no longer be retrieved. And write down what may run without approval, because data storage in Germany and AI processing in the EU are a foundation, but they do not replace the decision about what an agent may do without asking. You will find more background on our security page.

What You Can Take Away

When you face the question of whether a task belongs to a free agent or to a workflow, go through these questions in order:

1. 1.How often does the task occur in the same form? Several times a month and always the same: toward a workflow. Rarely, or different every time: toward free.
2. 2.Can I write the steps down today? If not, a workflow is premature. Start free and learn what the route looks like.
3. 3.What happens if the result is wrong? If it ends with a second attempt, freedom is defensible. If something goes outside, a fixed checkpoint belongs in front of it.
4. 4.Can I tell within minutes whether the result is right? If yes, check the end. If no, you need intermediate steps and a log.
5. 5.Which tools may the agent use? Reading can be generous. Writing only after asking or behind an approval.
6. 6.Who checks, and how long may it take? An approval without a named person and without a deadline is not an approval.
7. 7.When will I look at it again? Set a date on which you compare the log with the task.

Our position in one sentence: start with freedom where it costs nothing, and replace it with structure where a mistake gets expensive. The question is rarely free or workflow. Usually it is: which step needs freedom, and which one needs support? Whoever answers that for each task separately does not need a decision of principle.

If you would like to see how both feel in TheroAI, we are happy to show you on your own examples in a demo, without obligation.