AI agents
AI agent vs workflow: which one, and when to use both
In short
An AI workflow is a process written as steps that runs the same way every time something happens, such as a form being sent or a schedule coming round, with a model or agent doing only the steps that need judgement. An AI agent is a language model that decides its own steps as it goes, using knowledge and tools to finish a task a person gives it. Use a workflow when you can write the process down, an agent when the path depends on the request, and both together when a fixed process has steps that need reading and judgement.
An AI workflow runs the same steps every time something happens. An AI agent decides its own steps as it goes. Most real business processes need a bit of both: fixed steps where the process is known, and an agent where a step needs reading and judgement.
What is an AI workflow?
An AI workflow is a process written as steps. It starts on something (a schedule, an event in an app, another system calling it) and then does its steps in order, the same way every time. Some steps are plain actions, such as adding a row to a sheet or posting to Slack. Others ask a model or an agent for judgement, such as sorting a message or scoring a lead.
Because the steps are fixed, you can see in advance what a workflow will do, test it, and check each run afterwards.
What is an AI agent?
An AI agent is a language model that can use tools. A person gives it a question or a task, and the agent works out what to do: it searches knowledge, calls tools and decides the next step from what it found. Anthropic’s guide Building effective agents draws the same line: in a workflow, models and tools follow paths set in code; in an agent, the model directs its own process and tool use.
An agent handles requests nobody planned for. The cost is that each conversation can take a different path. For more, see AI agent examples.
AI agent vs workflow: the differences
| AI workflow | AI agent | |
|---|---|---|
| Who decides the steps | You, in advance | The agent, while it works |
| What starts it | An event: a schedule, an app event, a call from another system | A person asking, or another system calling it |
| Same input, same path | Yes | Not always |
| Best for | Processes you can write down | Questions and tasks that vary |
| How you check it | Test it, then read each run’s steps | Test conversations, then review answers |
| Where judgement comes from | The steps that ask a model or an agent | The whole run |
When should you use a workflow, an agent, or both?
Use a workflow when the same thing happens often and you can say what should happen each time:
- A website form is sent: sort it, skip spam, post leads to sales.
- An invoice is paid: thank the customer and log it.
- Friday comes: collect the numbers and send the weekly report.
Use an agent when a person asks and the path depends on the request:
- “How many days of parental leave do I get?” from the HR handbook.
- “Summarise what changed in this contract since last year.”
- “Find the open Jira issues that block the release and tell me who owns them.”
Use both when a fixed process has a step that needs reading and judgement. The workflow keeps the order, the record and the approvals; the agent does the reading:
- A new support e-mail arrives (workflow), the support agent drafts a reply from the help center (agent), a person approves, the reply is sent (workflow).
- A lead comes in (workflow), an agent scores it from the company’s notes (agent), the good ones go to Slack (workflow).
A simple rule: the more of a process you can write down, the more of it should be a workflow.
How do agents and workflows work together in promptev?
In promptev, both live in the same project, with the same knowledge, tools and approval rules.
- A workflow step can ask an agent. A step asks one of the project’s published agents, which uses its own knowledge and tools to read, judge or write. A step can also ask a model for an answer in a set shape, such as a score from 1 to 10, that is checked before the next step uses it.
- An agent can start a workflow. Tick the workflow in the agent’s Tools step. When a request is that workflow’s job, the agent starts it and tells the person the result.
- You describe a workflow in plain words. The workflow builder drafts the steps, draws them as a diagram, shows what one run costs, and lets you test it for real before you publish.
- People stay in charge. A workflow step that changes something can wait for a named approver, who decides by e-mail or in promptev. Agent tools have their own approval rules. See human in the loop AI.
- Every run on record. Each workflow run has its own page with the version it ran, the path it took and what each step got and gave back, and admins see every workflow run in the Activity log.
- For bigger jobs, a team of agents. An orchestration is a lead agent that directs other agents, for work that needs several of them.
Examples of agents and workflows together
| Process | Workflow part | Agent part | Person |
|---|---|---|---|
| Support e-mail | Starts on a new e-mail, sends the reply | Drafts the reply from the help center | Approves the reply |
| New lead | Starts on the website form, adds a row, posts to Slack | Scores the lead from 1 to 10 | None, internal only |
| Weekly report | Starts every Friday, saves and mails the report | Writes the summary from the numbers | None, internal only |
| Refund request | Starts on the form, checks the order, issues the refund | Reads the request and the order history | Answers “Refund $900 to Ana?” |
| New customer | An agent starts it when asked; it sends the welcome e-mail and tells the team in Slack | Notices the request is onboarding and starts the workflow | Approves the welcome e-mail |
For a longer list, see AI workflow examples.
Frequently asked questions
What is the difference between an AI agent and an AI workflow?
An AI workflow follows steps that were agreed in advance, in the same order every time something happens. An AI agent decides which steps to take while it works, based on the request, its knowledge and its tools. A workflow is easier to predict and check; an agent handles requests nobody planned for.
When should I use a workflow instead of an agent?
Use a workflow when the same thing happens often and you can write down what should happen each time: a form is sent, an invoice is paid, a new ticket arrives, Friday comes. Use an agent when a person asks something and the right steps depend on what they asked.
Can an AI agent and a workflow work together?
Yes. A workflow step can ask an agent to read, judge or write, such as scoring a lead or drafting a reply from the help center. An agent can start a workflow when a request matches that workflow's job, so a fixed process runs the same way however it was asked for.
Is an AI workflow safer than an AI agent?
A workflow is easier to check because its steps are fixed and every run takes a known path. Both still need controls: approvals before actions that send, change or delete, a cost limit on each run, and a record of what each run did.
Do I need to code to build an AI workflow?
Not always. Some tools have you connect boxes on a canvas. In promptev you describe the process in plain words, and the workflow builder drafts the steps, draws them as a diagram and lets you test it for real before you publish.

Faisal Saeed is Founder & CEO of promptev, the platform for sovereign AI agents for business: agents that answer from company data, see only what each person may see, and ask before they act.