The Workflow That Couldn’t Handle Surprises#

Here is the workflow we built first. A customer emails in. The system parses the subject line for keywords. “Billing” goes to Finance. “Bug” goes to Engineering. “Cancel” goes to Retention. Clean, fast, predictable.

Then reality showed up. A customer wrote: “I found a bug in my invoice and want to cancel if you can’t fix it.” The workflow matched “bug” first, routed to Engineering, and the customer waited three days for someone to tell them they were in the wrong queue. By then, they had already cancelled through their bank.

The workflow did exactly what we told it to do. That was the problem. We had assumed customer messages would fit neatly into buckets. They almost never do.

KEY INSIGHT: Workflows execute your instructions perfectly. Agents interpret your intentions. Knowing which you need depends on how predictable your inputs actually are, not how predictable you wish they were.

What an Agent Actually Does#

An agent reads context, makes decisions, and picks its own tools. You give it a goal (“route this customer to the right team”) rather than a script (“if subject contains X, send to Y”).

When we replaced our keyword-matching workflow with an agent, it handled that same multi-issue ticket by summarizing the customer’s concerns, identifying the primary intent (billing dispute), routing to Finance, and flagging the cancellation risk for Retention to follow up. One ticket, two departments notified, zero hardcoded rules.

What makes agents different:

• They decide on the fly. An agent encountering a new type of request does not crash. It reasons through it.
• They use tools dynamically. Need to search a knowledge base? Query a database? Call an API? The agent picks the tool based on what the situation requires.
• They produce variable outputs. The same input might yield slightly different responses because the agent weighs context, not just keywords.

That variability is both the superpower and the risk. You trade perfect consistency for the ability to handle situations nobody anticipated.

What a Workflow Actually Does#

A workflow is a pipeline. Data enters at step one, flows through predefined transformations, and exits at the end. Every run with the same input produces the same output. Full stop.

What makes workflows different:

• They are deterministic. Same input, same output, every single time.
• They are auditable. You can trace exactly which step processed which data and why.
• They scale cheaply. Running a workflow 10,000 times costs a fraction of running an agent 10,000 times.

Think about sending a welcome email when a new user signs up. There is no ambiguity. No context to interpret. No decision to make. A workflow handles this in milliseconds and never gets it wrong.

KEY INSIGHT: If you can write out every possible path on a whiteboard, use a workflow. If the paths depend on context you cannot fully predict, you need an agent.

Side by Side: Where Each One Wins#

Real examples make this clearer. We use workflows for invoice generation, deployment pipelines, and data backups. We use agents for code review triage, customer intent classification, and research summarization. The dividing line is always the same question: do we know all the possible inputs?

The Hybrid Approach That Actually Ships#

Most production systems end up using both. We learned this after over-engineering an agent to handle tasks that were perfectly suited for a simple workflow.

Our first version of a document processing system used an agent for everything, from extracting metadata to filing documents in the right folder. It worked, but it cost 15x more per document than it needed to. The extraction step was always the same: pull the date, author, and title from a structured PDF header. No judgment needed.

We rebuilt it as a workflow for extraction feeding into an agent for classification. The workflow pulled structured fields reliably and cheaply. The agent handled the messy part, deciding which project folder a document belonged to based on its content. Processing cost dropped by 80% and accuracy stayed the same.

KEY INSIGHT: Start every automation as a workflow. Promote individual steps to agents only when the workflow fails on real-world variability, not when you imagine it might.

Choosing Your Approach#

Reach for a workflow when:

• Your process has clear, repeatable steps that rarely change.
• You need guaranteed consistency and easy debugging.
• Speed and cost per execution matter more than flexibility.

Reach for an agent when:

• Inputs are unpredictable or require interpretation.
• The task involves choosing between multiple tools or strategies.
• You need context-sensitive reasoning, not just pattern matching.

Use both when:

• Predictable steps feed into unpredictable decisions (or vice versa).
• You want to keep costs low without sacrificing adaptability where it counts.

The Bottom Line#

Anthropic puts it well: “start simple and add complexity only when needed.” We have seen teams burn months building agent-powered systems for problems a 20-line workflow could have solved. We have also seen teams duct-tape workflows together for tasks that clearly needed an agent’s judgment.

The difference between a system that ships and one that stalls is almost always the same: someone asked “do we actually need an agent here?” before writing the first line of code.

KEY INSIGHT: The best AI systems are not the most sophisticated. They are the ones where every component, agent or workflow, earns its complexity.

References

• Anthropic’s Guide to Building Lean AI Agents
• How AI Agents Are Transforming Business - Forbes