AI workflow vs AI agent vs skill: what the job needs
What a prompt, skill, workflow, connector, agent, and custom software each mean, and how to pick the simplest one that does the job.
King & Company
In short
A workflow follows steps its builder set ahead of time, while an agent lets the model pick its own steps as it goes. A prompt, a skill, a connector, and custom software are the other four options a firm has. Choose the simplest one that does the job, and add complexity only when the work shows you need it.
The difference between an AI workflow and an AI agent is who decides the next step. In a workflow, the person who built it fixed the steps in advance, and in an agent the model chooses its own steps as the work unfolds. Those are two of six forms a piece of AI work can take, and a firm that is being pitched agents will spend its money better once it can name all six and say which one a given job calls for.
Why the vocabulary matters before you buy anything
If you run a brokerage team, an accounting firm, or a portfolio company, you have probably sat through a pitch with the word "agent" in it. The word covers a wide range of things. Sometimes it describes a system where the model plans and acts on its own, and sometimes it is a label on a scheduled script with one model call in the middle. Those two things have different costs, different failure modes, and different review needs, so the label alone tells you very little about what you would be buying.
A firm owner has more choices than workflow or agent. The same piece of work could be handled by a well-written prompt, a skill, a fixed workflow, a connector to another system, an agent, or a purpose-built application, and the right choice depends on the work.
Prompt, skill, workflow, connector, agent, software: one definition each
Prompt. A written request to the model, with the context and examples it needs. Anthropic's guide to building agents says that for many applications, "optimizing single LLM calls with retrieval and in-context examples is usually enough."
Skill. A packaged procedure. Anthropic's Help Center describes skills as "folders of instructions, scripts, and resources that Claude loads dynamically to improve performance on specialized tasks." A skill is where your firm's way of doing a task lives, along with the template it should fill in. We cover them in more depth in What are Claude skills?
Workflow. Anthropic defines workflows as "systems where LLMs and tools are orchestrated through predefined code paths." The steps and their order are set by the builder, and the model does a defined job at each one.
Connector. A link between the model and a system that holds your data. The common standard is the Model Context Protocol, which its own documentation describes as "an open-source standard for connecting AI applications to external systems" and compares to a USB-C port for AI applications. In plain English, MCP is the plug that lets an AI assistant read from and act in your CRM, your file store, or your database, and the same documentation lists both Claude and ChatGPT among the assistants that support it.
Agent. Anthropic defines agents as "systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks." The model decides what to do next, does it, looks at the result, and decides again.
Custom software. An application built for one job, with its own interface and usually its own database. A model may sit inside it, but the people using it see screens and buttons and may never write a prompt.
The ladder: start with the simplest form that does the job
We think of these six as a ladder, ordered by how much there is to build, test, and maintain. This ordering is our working view as builders, and it follows the advice in Anthropic's guide, which recommends "finding the simplest solution possible, and only increasing complexity when needed."
| Form | Use it when | What you take on |
|---|---|---|
| Prompt | The task is occasional and one person can describe it well | Almost nothing, since the prompt lives in someone's notes |
| Skill | The same procedure repeats and needs the firm's templates and standards | A folder of plain files that someone on the team owns and edits |
| Fixed workflow | The steps are known in advance and the output must look the same every time | A defined sequence to test and keep current as the inputs change |
| Connector or integration | The work stalls because the data is in another system | Credentials, permissions, and upkeep when the other system changes |
| Agent | The path cannot be known until the work is under way, and a wrong step is cheap to catch | A test set, a strong review step, and a higher cost per run |
| Custom software | Many people need the same interface, the work needs its own database, or the volume has outgrown a chat window | A codebase, hosting, and someone responsible for it |
The rows are not exclusive. A connector often sits under a skill or a workflow, and a finished system usually uses several of them. What the ladder gives you is a default: begin with the simplest form that could work and move to the next one when the work shows you it is needed.
When is a fixed workflow the right answer?
A fixed workflow fits when you can write the steps on a whiteboard before any work starts. Anthropic's guide says that workflows "offer predictability and consistency for well-defined tasks." Most recurring document work at a professional firm is well defined in this sense.
Take lease abstraction as an illustration. The steps are known: read the lease and its amendments, pull a fixed list of terms, note the page each one came from, fill in the firm's abstract template, and flag anything that conflicts. There is one real judgment step, which is deciding how an amendment changes the original term. That step can be handled by the model with a person checking it, and everything around it can stay fixed.
Recurring document work that gets described as needing an agent often has this shape. It is a known sequence with one or two points of judgment, and our view is that it is better built as a workflow with those points marked for review. We describe the build process in How to build an AI workflow your team will use every day.
When does an AI agent earn its cost?
An agent fits when nobody can predict the steps. Anthropic's guide recommends agents "for open-ended problems where it's difficult or impossible to predict the required number of steps," and it is direct about the price. Agentic systems "often trade latency and cost for better task performance," and "the autonomous nature of agents means higher costs, and the potential for compounding errors." The same guide recommends extensive testing in sandboxed environments, along with appropriate guardrails.
Two conditions should both hold before a firm chooses an agent. The path has to be unknowable in advance, as it is in open-ended research across many sources where each finding changes what to look at next. A wrong step also has to be cheap to catch and undo. Research that a person reads before using meets that test. Sending an email to a client or writing to a system of record does not, unless a person approves the action first.
The more freedom the system has, the more weight falls on the review step and the test set. We cover both in How to design the human review step in an AI workflow and in AI agent permissions and governance.
When is the real problem integration?
Sometimes the model already does the task well when someone pastes the right information in front of it, and the trouble is that the information lives somewhere else. The Claude Code documentation names this trigger plainly: when you "keep copying data from a browser tab Claude can't see," connect that system. A smarter model or a more autonomous one does not fix a data access problem.
Anthropic's September 2025 Economic Index report points the same way for businesses that use Claude through its API. It found that 77% of business uses of Claude through the API involve automation patterns, compared to about 50% for Claude.ai users, and it observed that deploying AI for complex tasks "might be constrained more by access to information than on underlying model capabilities."
A connector deserves restraint too. Anthropic's guidance on building tools for agents says, "More tools don't always lead to better outcomes," and recommends "building a few thoughtful tools targeting specific high-impact workflows." Connect the two or three systems the work depends on before connecting everything.
When should a firm build custom software?
Purpose-built software is the right choice in three situations. Many people need the same interface and should not each have to learn to prompt. The work needs its own database, with history, permissions, and reporting. Or the volume is high enough that doing the task one conversation at a time is the wrong shape for it.
Software carries the largest upkeep of the six forms, because it needs hosting, updates, and someone who can change the code. That cost is justified when the conditions above hold. When one person runs the task a few times a week, a skill or a workflow does the same job with much less to maintain.
How do the pieces combine in a real system?
A working system usually combines several of these forms. The most common pairing is a connector with a skill. The Help Center states the division of labor plainly: "MCP connects Claude to external services and data sources," and skills "provide procedural knowledge." The Claude Code documentation describes the same pattern, where MCP provides the connection and a skill teaches Claude how to use it well. A connector can reach your CRM, and the skill tells the model which fields your team fills in and how a record should read.
One distinction from that documentation matters to anyone responsible for risk. An instruction written into a skill "is a request, not a guarantee." The model reads it and usually follows it. If a rule has to hold every time, such as never sending anything outside the firm without approval, it belongs in a control that runs outside the model. In Claude Code that control is called a hook, and the documentation describes it as enforcement. Ask where each of your important rules lives, and have your compliance lead or IT owner confirm the answer for the tools your firm runs.
The last point is about upkeep. A system that nobody on the team can read is hard to keep current, whatever it cost to build. A skill is a set of plain files a team lead can open and edit. A workflow can be documented step by step. We prefer those forms where they do the job, because the client owns what we build and has to be able to change it after we leave.
Questions to ask a vendor who says you need an agent
- Which steps in this system are fixed, and at which steps does the model decide what happens next?
- What happens when the model takes a wrong step, and who sees it before a client does?
- What test set did you use, and can we run our own documents through it?
- Which of our rules are enforced by the system, and which are instructions the model is asked to follow?
- What does one run cost, and how long does it take?
- Could a fixed workflow or a skill produce the same output? If you tried that first, what fell short?
- When you are gone, what will our team be able to open, read, and change?
A vendor with a sound design will answer all seven without difficulty. If you would like a second opinion on which form a specific piece of your work calls for, you can get in touch.
Common questions
What is the difference between an AI workflow and an AI agent?
In a workflow, the person who built it decided the steps and their order ahead of time, and the model does a defined job at each step. In an agent, the model decides which step to take next and which tools to use while the work is under way. Workflows are more predictable, and agents are more flexible at a higher cost per run.
Is a Claude skill the same thing as an agent?
No, they are different things. A skill is a folder of instructions, scripts, and reference files that Claude loads when a task calls for it, so it carries your procedure and your templates. An agent is a way of running the model where it directs its own steps. An agent can use a skill, and so can a person working in an ordinary chat.
What is MCP and do I need it?
MCP, the Model Context Protocol, is an open standard for connecting AI applications to outside systems such as databases, file stores, and business software. You need it, or a connector built on it, when the work stalls because the data lives in a system the model cannot see. If your team is working from documents they upload by hand and that is fine, you can wait.
Which is cheaper to maintain, a workflow or an agent?
A fixed workflow is usually the cheaper one to maintain, because you can read the steps, test each one, and see where a failure happened. An agent can take a different path on different runs, so it needs a larger test set and closer review. Anthropic's own guidance notes that autonomy brings higher costs and the potential for compounding errors.
When should a firm build custom software instead of using skills and workflows?
Build software when many people need the same screen, when the work needs its own database and history, or when the volume is high enough that a chat window is the wrong place to do it. If one person runs the task a few times a week from a document, a skill or a workflow will usually do the job for far less upkeep.