There is a reason so many AI proposals arrive describing an agent.
Agents demo well. A system that decides its own next move looks like the future in a way that a form and three API calls does not. If you are selling the work, the agent is the easier thing to sell.
It is also, in most cases, the wrong answer. And the clearest statement of that comes from the company whose models the agents would be built on.
What Anthropic actually recommends
Their guidance on building agents opens with restraint rather than ambition:
When building applications with LLMs, we recommend finding the simplest solution possible, and only increasing complexity when needed.
They go further and name the trade you are making when you ignore that advice. Agentic systems “often trade latency and cost for better task performance”, and for a great many applications “optimizing single LLM calls with retrieval and in-context examples is usually enough”.
That is the model provider telling you to buy less than you were about to.
The four levels
Think of it as a ladder. Start at the bottom and stop at the first rung that solves the problem.
Level one is no model at all. Conditionals, a template, an integration, a form. If the decision is a rule somebody can write down, writing it down is cheaper, faster and correct every time. A surprising amount of what gets scoped as AI work belongs here.
Level two is a single call. One model call, with retrieval and good examples, sitting inside code you control. You decide what goes in. You decide what happens to what comes out. Most document extraction, most classification and most drafting jobs are level two, and they are boring to demo and excellent to run.
Level three is a workflow. Model calls at the points where judgement is needed, with your code orchestrating the sequence. The key detail is who holds the plan. In a workflow, your code does.
Level four is an agent. The model decides what to do next. You use this when the number of steps cannot be predicted and no path could sensibly be hardcoded.
Most business processes look like level three. Take a ten step process and inspect it properly. Perhaps three steps need judgement. The rest are conditionals and API calls wearing a disguise.
What level four actually costs
Anthropic are specific about the risk, and it is worth quoting to anyone asking you to build one:
The autonomous nature of agents means higher costs, and the potential for compounding errors. We recommend extensive testing in sandboxed environments, along with the appropriate guardrails.
Compounding errors is the part that matters, and it is a structural difference rather than a quality problem.
A workflow’s blast radius is one step. If step four goes wrong, step four is wrong, and the gate before step five catches it.
An agent’s blast radius is the whole run. Step four builds on step two, so a small error at step two arrives at step nine having grown. Nobody sees it happen, because the sequence was never written down anywhere a person could read.
There is a second cost that shows up later. A workflow can be tested. You can run the same fixed inputs through it after every change and compare the outputs to what they were last week. An agent that chooses its own path produces a different path each time, which makes “is this still working?” a much harder question to answer, and a much more expensive one.
Why the market points the other way
Nobody is being dishonest about this. The incentives just line up badly.
An agent is a better story in a meeting. It sounds like more value, so it justifies a larger number on the proposal. It also transfers the hard thinking from the supplier to the model, which is convenient when the supplier has not spent enough time understanding the process to know where the judgement actually sits.
Working out that seven of the ten steps are deterministic takes real effort. It is unglamorous, it makes the project smaller, and it produces something that works.
How to tell which level you need
Four questions, and you can ask them of any proposal in front of you.
Can you write down the steps in advance? If yes, you do not need an agent. You need code that runs those steps and calls a model at the two or three points where a judgement is required.
What happens when it gets something wrong halfway through? If the answer involves the system carrying on regardless, you are looking at compounding errors and you should ask what stops them.
How will we know next month that it still works? A supplier who cannot answer this concretely has not thought about testing, which is a bigger problem than the architecture. We wrote separately about why the acceptance criterion should be a number.
What does this buy that a simpler version does not? Agents trade cost and latency for capability on genuinely open-ended tasks. That trade is sometimes worth making. It should be named, not assumed.
The objection worth taking seriously
There is a fair counter-argument, and it deserves a hearing.
Building a workflow means somebody has to understand the process well enough to encode it. That is real work, and it goes stale when the process changes. An agent absorbs some of that change without anyone rewriting anything, which has genuine value in a business where the process shifts every quarter.
That is true, and it is the case for level four when it applies. It is also the reason to be honest about what you are buying: flexibility, paid for in cost, latency and testability. A supplier who presents an agent as strictly better than a workflow has not made that trade explicitly, and probably has not made it at all.
The short version
Start at the bottom of the ladder. Climb only when the rung you are on will not hold the problem.
For most small and mid-sized businesses that means a workflow with model calls at the judgement points, not an autonomous system planning its own route. It is cheaper to run, easier to test, and considerably easier to explain to the person who has to live with it.
If the proposal in front of you starts at level four without walking through the levels below it, that is worth a question before it is worth a signature.
Frequently asked questions
What is the difference between an AI workflow and an AI agent?
In a workflow, your code holds the plan and calls a model at specific points where judgement is needed. In an agent, the model decides what to do next. The practical consequence is testability: a workflow runs the same path every time and can be regression tested, while an agent chooses its path at runtime.
Do I need an AI agent for my business?
Usually not. Anthropic’s own guidance recommends the simplest solution that works, and notes that a single model call with retrieval and good examples is enough for many applications. Agents are for tasks where the number of steps cannot be predicted in advance.
Why do so many suppliers propose agents then?
They demo better and they justify larger projects. Determining that most of a process is deterministic takes more analysis than assuming a model can work it out, and it produces a smaller piece of work.
What are compounding errors?
An agent’s later steps build on its earlier ones, so a small mistake early can grow through the run rather than being caught. A workflow contains the damage to a single step, because your code checks the result before the next step begins.
Is a simpler system less capable?
For an open-ended task, sometimes. For the kind of repeatable business process most firms want automated, a workflow is usually more capable in the way that matters, because it is reliable enough to leave running and cheap enough to run often.
Flux Dynamics is a fractional CTO who builds. We work out where judgement actually sits in a process before proposing anything, which usually makes the build smaller than expected. Tell us what you are considering.