Agents vs workflows vs chatbots: which you need
How to tell a chatbot, a workflow and an agent apart, and a simple way to pick the cheapest one that does the job.
Updated 2026-11-19. "(source N)" refers to the numbered list at the end.
Three different tools
People use the words chatbot, workflow and agent as if they mean the same thing. They do not, and the difference matters for cost, risk and how much time you will spend fixing things.
Here is the plain version.
- A chatbot answers. You ask, it replies. It may look things up, but it does not carry out a task in your other systems.
- A workflow follows steps you wrote. Some steps may call a language model, for example to summarise an email, but the order of steps is fixed in code (source 1).
- An agent decides its own steps. The model looks at the goal, picks tools, checks results and keeps going until done (source 1).
Why simpler usually wins
Anthropic, which builds agents for a living, tells developers to begin with the simplest option and add complexity only when it clearly helps. Often a single well-written model call, with the right reference material pasted in, is enough (source 1).
The reason is cost and risk. Agents use more tokens because they loop, and each loop is another chance to go wrong. Anthropic notes that mistakes can compound over many steps (source 1). A workflow, by contrast, does the same thing every time, which makes it easy to test.
OpenAI's guide says the same from the other side. Before you build an agent, check that the task really needs one. If it does not, a fixed, rule-based approach may be all you need (source 2).
Workflow patterns that cover most needs
Anthropic describes a handful of workflow shapes that solve many problems without full autonomy (source 1). Each one keeps you in charge of the order of steps.
- Chaining: one model call feeds the next. Example: draft a product description, then check it against your style rules.
- Routing: a model sorts the input, then code sends it down the right path. Example: billing emails go to one prompt, returns to another.
- Parallel calls: run several checks at once and combine them. Example: three reviewers score a contract clause and you take the majority.
- Orchestrator and workers: one model splits a job into parts for other calls. This is the step just before a true agent.
- Generate and critique: one call writes, another grades, and they loop until the grade passes.
When an agent earns its keep
OpenAI points to three kinds of work where agents beat rules (source 2). First, judgment calls with lots of exceptions, such as whether to approve a refund. Second, rule books that have grown so big no one wants to touch them, such as vendor security reviews. Third, work buried in messy text, such as reading an insurance claim.
Real usage lines up with this. In LangChain's practitioner survey, the most common agent jobs were customer service at 26.5% and research and data analysis at 24.4% (source 3). Both involve open-ended text and a lot of small decisions.
Anthropic's own usage data shows another pattern. When businesses call Claude through the API, the work is far more directive, meaning the model is told to do a job with little back-and-forth. Two API uses that at least doubled in share over three months were sales outreach and market monitoring (source 4).
A quick way to choose
Answer these in order and stop at the first yes.
- Is this one question with one answer, like tag this ticket or extract this date? Use a single model call. Our decision models article covers this case.
- Do people mostly need answers from your documents? Use a chatbot grounded in those documents.
- Can you write the steps on a whiteboard and they rarely change? Build a workflow.
- Do the steps change case by case, and does the task need several tools? Now consider an agent.
Mixing them is normal
Most good systems blend all three. A shop might use a chatbot on its website, a workflow to file each new order, and a small agent that handles return requests that do not fit the usual rules.
Start with the workflow and only hand the hard branch to an agent. That keeps the agent's job narrow, which makes it easier to test and cheaper to run.
Whatever you pick, plan an exit. OpenAI recommends that agents hand control back to a person when they keep failing or when an action is risky, such as a large refund or a cancelled order (source 2). A workflow can include the same check as a fixed step.
Cost and speed in practice
A chatbot reply might use a few thousand tokens. An agent that loops ten times, each time re-reading its instructions and tool list, can use ten or twenty times that. Speed follows the same pattern: each loop is another round trip to the model.
That is why the question is not which is best but which is enough. Pick the smallest tool that meets your quality bar, measure it, and upgrade only when the numbers say you should.
Sources
Written by Agents Lookup in our own words. Facts come from these pages:
- 1. Anthropic: Building effective agents: www.anthropic.com/engineering/building-effective-agents
- 2. OpenAI: A practical guide to building agents: cdn.openai.com/business-guides-and-resources/a-practical-guide-to-building-agents.pdf
- 3. LangChain: State of Agent Engineering: www.langchain.com/state-of-agent-engineering
- 4. Anthropic: Economic Index report, Learning curves (March 2026): www.anthropic.com/research/economic-index-march-2026-report