Which should a given process use?

Ask what should happen to a case nobody anticipated. If the right answer is to stop and tell somebody, a fixed path is correct and an agent will paper over the gap. If the right answer is to attempt it sensibly, the deciding kind earns its unpredictability. Most real processes contain both kinds of step.

Why is this harder to budget for?

Because the number of steps is not known before the run. A fixed path costs the same every time and can be multiplied by volume. One that decides as it goes takes as many turns as the case requires, so the bill follows the difficulty of the inputs rather than their count.

Does it need supervision?

More than a fixed path, and it usually gets less, because these run on triggers when nobody is present. That combination is why the limits and the autonomy setting carry more weight here: they are the only things in the room at the moment the decision is made.

Glossary

Agentic workflow

An agentic workflow lets the model decide what happens next instead of following a path somebody drew, which handles variation well and gives up being predictable.

In plain terms

An ordinary automated process does the same steps every time: when this happens, do that, then that. An agentic one is told what the outcome should be and works out the steps itself, which may differ from run to run. The first is reliable and brittle. The second is adaptable and unpredictable, and those are the same property described twice.

Why it matters

Because it is the choice that decides whether a process is trustworthy, and it is frequently made by whoever happened to build the thing. Work that must produce identical results every time and work that must cope with cases nobody listed want opposite answers, and a tool bought for one is regularly pointed at the other.

How it works

The difference is who chooses the next step. A fixed path was decided in advance by a person and runs the same way on every record. An agentic one is decided during the run, from what the model finds, which is why the same input can produce a different route on a different day without anything having been changed.

Variation is what it is for, and the reason is straightforward. Ordinary automation fails on the case nobody anticipated, because there is no branch for it, and every exception has to be added by hand. Something deciding as it goes can attempt a case that was never listed, which is precisely the work that made the exception queue long in the first place.

The cost is that you cannot say what it will do. Two runs on similar inputs may take different routes, so testing establishes that it worked on those cases rather than that it works. Vendors are unusually direct about this: one platform notes it is less suited to deterministic high-volume pipelines where the same steps must execute identically every time.

Triggers rather than requests is what makes it a workflow rather than a conversation. These run on events, on a schedule or on a signal from another system, which means nobody is present when the decision is made. That is the whole benefit and it is also why the autonomy setting and the limits matter more here than in anything a person is watching.

The two are increasingly sold together, and the useful arrangement usually mixes them. Automation platforms now carry AI steps as first-class components inside otherwise fixed scenarios, which lets the deterministic parts stay deterministic and confines the judgement to the step that genuinely needs it. That is a more defensible shape than making the whole process agentic because one part of it was awkward.

The case nobody listed

Framing the choice around the unanticipated case cuts through most of the argument, because it replaces a question about technology with a question about consequences. In a payroll run, a case nobody listed should stop the process and summon a person, and an agent that handles it gracefully has done something worse than failing: it has produced a plausible outcome nobody reviewed. In a support queue, the same graceful handling is the entire point, since the alternative is a customer waiting for somebody to write a branch that will be used twice. The tools have converged on letting both live in one process for exactly this reason, which is why the practical version of this decision is rarely made once for a whole system. It is made per step, and the steps where it matters are usually few and obvious once somebody asks the question.

Seen in the wild

Agents set up in plain language that run on triggers rather than waiting to be asked, whose documentation says deterministic high-volume pipelines fit classic workflow tools better.
Lindy
A visual automation platform where agents live inside the scenario builder as reusable components, with their reasoning and tool calls visible step by step.
Make
A canvas where AI operations are first-class nodes rather than bolted-on steps, suiting data-processing jobs too AI-centric for classic automation.
Gumloop

Common misconceptions

People assume

Agentic is the upgrade and fixed paths are the old way.

In fact

They solve different problems and neither is a stage of the other. Work that must be identical every time is better served by a path somebody drew, permanently. The deciding kind earns its place where the cases vary more than anybody can enumerate, which is a property of the work rather than of how modern the tool is.

People assume

It can be tested the way an automation is tested.

In fact

Testing shows it handled those cases, not that it handles the class. Because the route is decided during the run, a passing test is weaker evidence than the same test on a fixed path, and the gap widens as the input varies. That is not a defect to be engineered away; it is what deciding at runtime means.

Telling them apart

Agentic workflow vs Workflow

Agentic workflow

The next step is decided during the run.

Workflow

The steps were drawn before anything ran.

Ask whether an unlisted case should be attempted or refused. That answer picks the shape, and it belongs to the work rather than to the tool.

Questions

Which should a given process use?: Ask what should happen to a case nobody anticipated. If the right answer is to stop and tell somebody, a fixed path is correct and an agent will paper over the gap. If the right answer is to attempt it sensibly, the deciding kind earns its unpredictability. Most real processes contain both kinds of step.
Can they be combined?: That is now the common shape, and usually the more defensible one. Automation platforms carry AI steps inside otherwise fixed scenarios, so the deterministic parts stay deterministic and the judgement is confined to the step that needs it. Making a whole process agentic because one step was awkward is the pattern worth avoiding.
Why is this harder to budget for?: Because the number of steps is not known before the run. A fixed path costs the same every time and can be multiplied by volume. One that decides as it goes takes as many turns as the case requires, so the bill follows the difficulty of the inputs rather than their count.
Does it need supervision?: More than a fixed path, and it usually gets less, because these run on triggers when nobody is present. That combination is why the limits and the autonomy setting carry more weight here: they are the only things in the room at the moment the decision is made.

Key takeaways

The difference is who chooses the next step, and when.
Adaptable and unpredictable are the same property described twice.
Testing shows it handled those cases, not that it handles the class.
It runs on triggers, so nobody is present when the decision is made.
Mixing the two, with judgement confined to the step that needs it, is the defensible shape.