First it was agents, now it’s agentic
Agentic AI, agentic workflows, agentic systems. The term shows up at every conference and in every vendor demo. And just like with agents, the hype hides a fairly concrete technique.
In our two-part series on AI agents we explained what an AI agent is, an LLM that calls tools, and when to use them and when not. An agentic workflow is the logical next step: how you connect several of those steps into a system that finishes a real task.
From a single agent to a working system
A single LLM call gives you an answer. An agent adds tools, so it can also do something. An agentic workflow adds structure: multiple steps, multiple tools and checks in between, all aimed at finishing a task.
The building block underneath is always the same: a model with memory, access to data and a set of tools. Put a few of those blocks in sequence, with logic in between, and you have a workflow.
A single LLM call
You ask a question and get text back. Fine for a summary or a translation. One step, done.
An agent
The same model, now with tools. It fetches data, sends an email or creates a ticket. A step becomes an action in your real systems.
An agentic workflow
Multiple steps in sequence, with tools and checks in between. Classify, fetch, generate, check. The system handles a whole task from start to finish.
A single agent does a step. An agentic workflow finishes a task. The difference is the structure you build around it.
Workflow or agent? That distinction decides your reliability
Within agentic systems, two approaches run side by side, and the distinction matters more than it sounds.
In a workflow, the LLM steps follow paths you define in code in advance. First classify, then the right tool, then a check. You know exactly which route the data takes.
In an agent, the model decides for itself, while running, which steps it takes and in what order. You give it a goal and a set of tools, and the agent picks its own way to the result.
Workflow: you define the paths
The steps are fixed in code. Predictable, testable, easy to monitor. You know on every run what happens and where it stalls if it stalls.
Agent: the model defines the paths
Flexible and powerful for open tasks. The trade-off is that you give up predictability, because each run can take a different path.
For most business processes a defined workflow is the better choice, precisely because you can verify and trust it. You reach for a truly autonomous agent the moment you cannot map the route in advance, for example with open research where you don’t know up front which sources are relevant.
We wrote in part 2 about the red button problem: an agent does what you say, even when that is something other than what you meant. A defined workflow suffers far less from this, because the space the model is allowed to move in is small and controlled.
The more freedom you give the model, the less you know in advance what it will do. For most business processes, predictability is worth more than freedom.
The five patterns you build agentic workflows from
Agentic workflows look complex. Under the hood they are built from a handful of patterns that keep coming back. Once you recognize them, you see through any agentic pitch.
Prompt chaining
A task split into fixed steps, where each LLM call builds on the previous one. Write a piece, check its tone, then translate it. Between the steps you put a check that stops the moment something is off.
Routing
A first step classifies the input and sends it to the right follow-up. Incoming support questions go to the route for billing, technical or sales. Each route is optimized for its own type of question.
Parallelization
Running several subtasks at once and combining the results. Review a contract on five risk points at the same time. Or ask the same question a few times and let the answers vote for a more reliable result.
Orchestrator and workers
A lead model breaks a large task into pieces, hands them to specialized workers and merges the results. Strong for research that spans many sources at once.
Evaluator and optimizer
One model produces something, a second judges it against clear criteria and sends it back for improvement. That loop runs until the result passes the check. Ideal for translations or copy where quality matters.
In practice you combine these patterns. A support system first routes the incoming question, then chains a few steps, and lets an evaluator check the result before an answer goes out.
The hard part is the system around it
You can rebuild the patterns above in an afternoon with n8n, Make or a few lines of code. That is the easy part. The hard part starts the moment the workflow has to run on real data every day, with write access to your systems, while the models underneath keep changing. Four layers around the model make the difference then, starting with the harness.
The harness
The system around the model: what the workflow may do, which tools it has and what happens when a step fails. This is where you set the boundaries the model is allowed to move in.
Evals
Automated tests that judge each step. They warn you when a model update quietly changes the outcome. Without evals you notice a regression only when a customer complains.
Monitoring
Visibility into what happens in production: which paths get chosen, where runs stall, what it costs. A workflow you cannot see is a workflow you cannot trust.
Guardrails
Checks that keep the workflow within bounds. A human who signs off on irreversible actions, hard limits on what a step may do.
Anyone can build an agentic demo. An agentic workflow that runs reliably in production for months, that is engineering.
This is exactly the layer we build for clients: AI agents and automation that keeps working after the demo, day in and day out.
When an agentic workflow is worth it
An agentic workflow is powerful and, for many tasks, also overkill. Here is the rule of thumb. Does a step genuinely need language understanding or a judgment? Then an LLM fits. Is it about moving data from A to B by fixed rules? Then a plain automation is faster, cheaper and more stable.
Choose a plain automation
Fixed rules, no judgment needed. A file copied every night, an alert on a new order. A Zapier flow or a cronjob is enough.
Choose an agentic workflow
The task involves language understanding, classification or a judgment. Understanding and answering incoming emails, summarizing documents, ordering unstructured data.
Always start as simple as possible. You add an extra step or more autonomy only once it pays for itself in a better result. Every step you add is also one that can fail.
Conclusion: define your paths, build the control around them
Agentic workflows become clear once you look past the term. They are multiple AI steps, structured so the whole finishes a real task, in a way you can build on.
Workflow: you define the paths
Fixed steps in code. Predictable and testable. The right choice for most business processes.
Agent: the model defines the paths
Flexible for open tasks, at the cost of predictability. Use it where you genuinely cannot fix the route in advance.
The value is in the system around it
Harness, evals and monitoring make the difference between a demo and something that holds up in production.
Start simple, define your paths and build the control around them before you give away more autonomy. Working in a demo and holding up in production are two different things.
What is an agentic workflow?
What is the difference between an agentic workflow and an AI agent?
What is the difference between an agentic workflow and plain automation?
What types of agentic workflows are there?
When do I need an agentic workflow?
An agentic workflow that holds up in production?
We build the layer around the model: the harness, the evals and the monitoring that keep a workflow reliable while your systems and the models change. From first setup to ongoing management.
