Choose a workflow when the steps and boundaries can be defined in advance. Consider an AI agent when the task needs flexible investigation and the system must choose its next step. Begin with the smallest scope you can evaluate.
An automation can use AI without becoming an autonomous agent. A fixed process might ask a model to summarize a message, validate the response and send a draft to a person. The surrounding application still controls the sequence.
This distinction matters when a business is deciding what to build, who remains responsible and how to test it. The useful question is how much freedom a particular task needs.
01 / Compare who chooses the next step
In a defined workflow, the application decides the path. A form arrives, required fields are checked, a model prepares a summary and a reviewer accepts or corrects it. The model can interpret the message without deciding which system to change.
In an agentic system, the model has more influence over the sequence. It may choose a tool, inspect the result and decide which action to attempt next. That flexibility can help with a task whose steps vary, but it also expands what needs testing.
Anthropic’s engineering explanation distinguishes predefined orchestration from model-directed processes and recommends adding complexity only when it is useful. Terminology varies across vendors; inspect the behavior behind the label.
| Decision | Defined workflow | AI agent |
|---|---|---|
| Next step | Application rules select it | Model helps select it |
| Useful starting task | Repeatable process with known branches | Investigation with variable steps |
| Testing focus | Inputs, outputs and each branch | Tool choices, action sequences and stopping |
| Human role | Review selected outputs and exceptions | Set boundaries and review consequential actions |
02 / Use a workflow for repeatable work
Consider a new customer request. If a form already records the service type and location, ordinary rules may assign it. If the customer describes the request in free text, an AI step can suggest a category while the application checks the suggested value.
That process has defined inputs, a small set of categories and a clear destination. A reviewer can compare the suggestion with the original request. Ambiguous cases can go to a person instead of expanding the model’s permission to improvise.
Other candidates include preparing a project brief, assembling an internal report from approved inputs or drafting meeting actions. Start with the first-project scorecard to check ownership, data and recovery before choosing an architecture.
03 / Consider an agent for variable investigation
A different task might ask a system to investigate why a non-sensitive internal build failed. The relevant files and checks can change from case to case. A system that can inspect a log, select a diagnostic tool and use the result to choose another check may be useful.
That does not mean it should automatically deploy a fix. Investigation and permission to change production are separate decisions. A first version can gather evidence and propose an action for review.
Google Cloud’s architecture guidance discusses patterns for open-ended and multi-step tasks. Treat those patterns as options to evaluate against a real task, rather than a reason to introduce several agents into a simple process.
Before selecting an agent, identify a representative set of cases. Explain what flexible step selection would add, how you would recognize an incorrect tool choice and where the process must stop.
04 / Define the permissions before the tools
List what the system may read, what it may propose and what it may change. Give it the access needed for the chosen task. A research task may need read access while a record update needs a separate write permission and validation.
Put approval at the action that matters. A person should see the proposed recipient and message before sending, or the exact record change before accepting it. A vague approval of “the plan” is difficult to apply when the system later chooses a different route.
Set a maximum run length or budget, a timeout and a destination for unresolved cases. Decide how the operator can stop a run. Preserve enough source evidence and action history to understand what happened.
Retrieved pages, documents and tool responses should be treated as task data. Their text should not be allowed to grant new permissions or redirect the system into unrelated actions. This is especially relevant when the task reads external material.
05 / Test the path as well as the answer
For a workflow, test every defined branch. For an agent, also inspect the sequence of choices. A correct final answer can conceal unnecessary tool calls, access outside the intended scope or a retry that creates duplicate work.
- Give it a complete, ordinary task and check the output.
- Remove a required input and check whether it asks or stops.
- Return conflicting information and inspect how it handles uncertainty.
- Make a tool unavailable and confirm the fallback.
- Repeat a request and check duplicate handling.
- Present an instruction inside retrieved material and confirm it cannot expand authority.
Use approved test material. Record completion quality, review effort, tool usage and unresolved cases. Compare the simplest workflow that could handle the task with the more flexible design. Expand autonomy only when the evidence shows a benefit worth its operating burden.
06 / Write the decision in plain language
We will use a fixed intake workflow with one AI classification step. It can suggest a service category, but a person handles ambiguous requests and approves any customer reply. We will reconsider an agent only if variable investigation becomes necessary.
This is an illustrative decision, not a description of a deployed client system. A short statement like this gives the builder and operator a shared boundary they can test.
Include the maintenance owner and a review date. If the process changes, update the decision and the test cases together. The automation cost guide helps account for the review, usage and maintenance work each option introduces.
Common questions
Can a workflow contain several AI steps?
Yes. The application can control a sequence of model calls, checks and approvals. The number of calls alone does not establish how autonomous the system is.
Does a business need multiple agents?
Only if distinct responsibilities and coordination solve an observed problem. Multiple agents introduce handoffs and additional behavior to evaluate.
Can an agent run without human review?
Some bounded tasks may justify it after evaluation. The decision depends on the allowed actions, consequences, recovery options and demonstrated reliability. Keep consequential actions explicitly controlled.
Sources & further reading
- Anthropic: Building effective agents. Published December 2024; used for architecture concepts, not current tool pricing.
- Google Cloud: Choose a design pattern for your agentic AI system.
Examples and evaluation prompts are original planning aids. Explore William’s selected work for business context.