The word "agent" now covers a chatbot on a website, a script that sorts email and a research tool that runs for an hour without supervision. When one word means all of those, it has stopped being useful for making decisions.
There is a real distinction underneath it. It is simple, it is not about how advanced a system is, and it decides most of what a business should care about: cost, risk and whether the thing can be tested.
Two definitions
The clearest statement of the difference comes from a 2024 engineering note by Anthropic, which builds one of the major families of language models.
A workflow is a system in which models and tools are orchestrated through code paths defined in advance. A designer wrote the steps. The model does the work inside each step.
An agent is a system in which the model directs its own process. It decides what to do next, which tools to use and when the task is finished.
The difference is who controls the sequence. In a workflow, the designer does. In an agent, the model does.
The agent pattern was set out by Shunyu Yao and colleagues under the name ReAct, for reasoning and acting, first posted in 2022 and published at a conference in 2023. The model thinks about what to do, takes an action such as searching or calling a tool, looks at the result and repeats the loop until it judges the task done.
What a workflow looks like
Take an invoice arriving by email.
- Read the attachment and extract the supplier, the date, the amount and the line items.
- Match the supplier against the list in the accounting system.
- Compare the amount with the purchase order.
- If everything matches, prepare the entry. If not, send it to a person with the difference marked.
- After approval, post the entry and file the document.
The order never changes. A model is used at step one, to read a document in any layout, and perhaps at step four, to write a clear note about what did not match. Everything else is ordinary software.
This is not a lesser use of AI. Reading a messy document correctly is the hard part, and it is the part that could not be automated before. The fixed sequence around it is what makes the result dependable.
What an agent looks like
Now take a different request: "Find out why last month's shipping costs were higher than usual."
Nobody can write the steps for that in advance. The first query might show that one carrier's charges rose. That suggests checking whether rates changed or volumes did. The answer to that decides the next question. The number of steps depends on what turns up.
An agent suits this. It needs room to choose.
The trade
Predictability. A workflow does the same thing every time. An agent may take a different route on two runs of the same request. For exploration that is acceptable. For posting entries to a ledger it is not.
Cost and speed. Each decision an agent makes is another call to a model. A task a workflow finishes in two calls may take an agent twenty. That is slower and costs more.
Compounding error. A small chance of a wrong turn at each step becomes a meaningful chance across many steps. A workflow has few steps and fixed ones. An agent on a long task has many.
Testing. A workflow can be tested like any software: this input at step two should give that output. An agent can only be tested on outcomes, across many runs, because the path is not fixed. It can be done, but it is more work and gives weaker guarantees.
Reach. This is where the agent wins. When the task cannot be specified in advance, a workflow cannot do it at all.
Most business work is a workflow
Look at the processes a business wants help with: answering the phone and booking the appointment, taking in documents and filing them, following up on quotes, chasing missing paperwork, producing the weekly report.
Every one has known steps. What varies is the content: the caller's words, the layout of the document, the wording of the reply. A model handles the variation inside each step. The sequence stays where a person put it.
Describing these as agents is common. Building them as agents is usually a mistake. It hands the model a decision, what to do next, that the business already knows the answer to.
Where an agent earns its place
There are jobs where the steps cannot be known ahead of time.
- Research across many documents, where each finding decides where to look next.
- Troubleshooting, where the path branches on what each check reveals.
- Preparing a first draft of something that needs material gathered from several systems.
Even here, the freedom should be bounded. A dependable agent has a short list of tools it may use and no others. It can read widely and write narrowly. It has a limit on steps and on spend. It keeps a log of what it did and why. And anything with consequences, such as sending, paying or deleting, waits for a person.
In practice the best designs are mixed. A fixed workflow runs the process, and at one step, where open-ended work is needed, it hands a bounded task to an agent and takes the result back.
Five questions to ask
Before deciding what a process needs, ask these.
- Can the steps be written down? If yes, it is a workflow.
- What happens if it takes a wrong step? If the answer is "a wrong draft", there is room for freedom. If it is "a wrong payment", there is not.
- Can each action be undone? Reading is safe. Writing to a system of record needs a gate.
- How would it be tested? If nobody can say what a correct run looks like, it is not ready to build.
- Who approves? Every design should have an answer, and the answer should be a person's name.
How we choose
The Anthropic note makes a recommendation we agree with: find the simplest solution that works, and add complexity only when it is needed. Often that means not building an agent.
In our work the order of preference is the same. A single well-prepared model call if that does the job. A workflow when the job has steps. An agent for the part of the job that cannot be specified, with limits around it. Each step up buys flexibility and costs predictability, so each one has to be justified by the task.
The question to ask of any proposal is not whether it uses agents. It is whether the system does the same right thing on an ordinary Tuesday as it did in the demonstration. Our approach page explains how we test for that before anything goes live, and approvals and routing covers the gate that sits in front of consequential actions.
References
- Schluntz, E., and Zhang, B. (2024). Building effective agents. Anthropic. anthropic.com
- Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., and Cao, Y. (2023). ReAct: synergizing reasoning and acting in language models. International Conference on Learning Representations. arXiv
