Workflow or Agent? Enterprise architecture choices based on Anthropic's definition

According to Anthropic's definition in "Building Effective Agents," the distinction is straightforward: Workflow uses predetermined execution paths to coordinate LLMs and tools. Agent allows the LLM to decide its own flow and tool usage, controlling how to complete tasks independently. One key difference, in Workflow, you decide the path; in Agent, the model decides the path in real time.

This definition might seem abstract, but it determines whether your system fails at 3 AM, whether costs grow uncontrollably, and whether you can trace what went wrong. Most enterprise selection debates simply reduce to not drawing this line clearly.

Workflow versus Agent: it's not about being smarter, it's about who has control

People often ask: "Is what you built for us a Workflow or an Agent?" Many assume Agent is the premium option, the more advanced choice, while Workflow is basic. That's incorrect.

The real difference is control. In Workflow, engineers hard-code each step and what comes next. The LLM only handles what it's good at on certain nodes: classifying, extracting, rewriting. The flow is fixed. Agent works differently: you give it a goal and a set of tools. It decides which tools to call, how many times, and when it's done. The path unfolds during execution.

Anthropic introduces another term: Agentic System. This is the umbrella category for both Workflow and Agent. LLM autonomy exists on a spectrum, not as a binary choice. What organizations actually choose isn't "do we want Agent," but rather "where on this spectrum should each step sit?"

When to use each approach

Apply this table step by step to each part of your system. Real systems typically use both: Workflow controls critical decisions at the outer layer, while a smaller Agent component handles only the specific nodes where actual flexibility matters.

How we learned this the hard way

A manufacturing client last year wanted to automate work order dispatch. The team was enthusiastic about the first version and built a fully autonomous Agent: feed it a fault description, let it search manuals, determine parts, create a dispatch ticket, notify the technician. The initial demo was impressive.

Two weeks after launch, problems appeared. The same failure type produced different results: one day the Agent searched three manuals, the next time seven. Token costs varied by a factor of three. Worse, it once assigned a high-priority repair to an intern because it reasoned this would be faster. When we tried to understand why, there was no readable decision trail.

We rebuilt the system from scratch. Dispatch decisions, errors here create liability, returned to Workflow: fixed rules plus a single LLM classification step. Only the which manuals to check and how to summarize part, where actual variability matters, retained Agent functionality. Costs became predictable again. Audits worked. Usage rose from 20% at launch to 70% in three months.

The principle we learned is straightforward: if Workflow solves the problem, use Workflow. Anthropic makes the same point in the same article, find the simplest solution first. Only tackle Agent's complexity and unpredictability when a task requires dynamic decisions and flexibility. Agent isn't the default; it's the exception.

Architecture decisions: three questions

When starting a project, we ask three things.

Can you map the steps of this stage on a whiteboard and be finished? If yes, it's Workflow. If no, if input variations create endless branches, then consider Agent.

What happens if this fails? High cost plus regulatory exposure? Choose predictability and traceability over flexibility. Don't let the model make this decision independently.

Can you see every decision it makes? If you're using Agent, you must build decision logs, tool call limits, and human review checkpoints at the same time. Without these, Agent autonomy masks risk.

Most organizations need a hybrid system: Workflow as the backbone, with Agent embedded in a few nodes where it adds value, not a fully autonomous system making all its own decisions.

This is how we approach agentic workflows at Tenten: map each stage according to Anthropic's definition, assign determinism to Workflow, reserve flexibility for genuine needs, deploy engineers to monitor it from prototype through production. Systems that work are common. The real measure is staying deployed and actively used.