AI agents vs workflows: when to use which in 2026

A workflow is a fixed sequence of steps with a predictable branch structure. The model may write the copy, classify the ticket, or extract the invoice fields, but the graph of what happens next is defined by you, not the model. Workflows are legible, testable, and cheap to run.

An agent is a loop. The model chooses the next action from a set of tools, observes the result, and chooses again until it decides the task is done or a stop condition fires. Agents are powerful when the branch structure is genuinely open, and dangerous when it is not, because they will happily invent branches that do not exist in your business.

The rule of thumb we use with clients: if a competent junior on your team could write down the steps on a whiteboard in ten minutes, it is a workflow. If they would need to say 'it depends' more than twice, it is a candidate for an agent, and even then only after you prove the workflow version is impossible.

1. How many distinct end states does the task have? Under five, workflow. Over twenty, agent. In between, workflow with an escape hatch.

2. What is the cost of a wrong action? If a bad output costs a refund, a workflow with human review beats an agent every time. If a bad output costs a rewritten email, the agent's speed advantage compounds.

3. How stable is the environment? Agents assume tools behave consistently. If your CRM API changes quarterly, agents will silently drift. Workflows fail loudly, which is what you want.

4. Who owns the failure? A workflow failure has an obvious owner: the step that broke. An agent failure requires a trace viewer and a patient engineer. If you do not have both, do not ship the agent.

5. Can you write the eval in one afternoon? If yes, workflow is enough. If the eval itself needs a rubric and a review board, you are in agent territory, and you should budget accordingly.

Anthropic's own guidance on building effective agents (published 2024) recommends starting with the simplest composition that works and only reaching for autonomous loops when the task genuinely requires them. LangChain's 2025 State of Agents survey found that the median production agent runs fewer than four tool calls per session and completes in under thirty seconds. In other words, most production 'agents' are workflows with a small planning step at the top.

This matches what we see in engagements. The systems that survive contact with real users tend to be workflows dressed up with a light routing layer, not the sprawling autonomous constructs the demos suggest. Simplicity ships. Sprawl gets rewritten.

Step one: write the task on one line, in the operator's words, not the vendor's. If it contains 'and', split it into two tasks and start over.

Step two: draw the ideal path on a whiteboard, then draw the three most common exceptions. If the exceptions fit on the same board, ship a workflow. If they need a second board, promote to an agent with a hard cap on tool calls.

Step three: define the eval before you write the prompt. If the eval is a checklist, workflow. If the eval is a rubric with a human tie-breaker, agent, and staff the tie-breaker before launch.

Step four: measure cost per successful task, not cost per call. Agents look expensive on cost per call and often win on cost per successful task. Workflows look cheap on cost per call and sometimes lose on cost per successful task because they need retries. Track the right number.

Roughly 80% of what we ship is workflow with a small routing layer at the top. The remaining 20% are true agent loops, always with tool-call caps, human review on high-stakes actions, and traces retained for at least ninety days. This ratio is not ideological. It is what survives the ninety-day review with process owners who have to live with the system.

If you want to see how this looks inside one interface, our note on how one connected AI system beats twenty disconnected tools walks through the control plane. And if you are still deciding whether to build or buy the underlying architecture, the guide to the best AI automation companies for growing teams frames the trade-offs cleanly.