Agentic Workflows: Definition, Patterns and Examples

An agentic workflow is an AI workflow whose overall path is fixed in advance, but in which some steps let the model decide: which branch a case goes to, how many subtasks a request is split into, which tool to call, or whether a draft is good enough to move on. It sits between the plain AI workflow, where the model only carries out defined steps, and the autonomous agent, which chooses its whole approach itself. For most business processes it is the sensible middle: enough flexibility to absorb varied cases, enough structure to test, predict cost and stay in control.

AI workflow, agentic workflow, agent: where is the line?

The vocabulary is loose, including among vendors. The European Commission notes in its FAQ on AI agents that these terms are often used inconsistently. The useful test is simple: who decides the next step?

	AI workflow	Agentic workflow	Autonomous agent
Path	Fixed end to end	Fixed in outline	Chosen by the model
Decisions by the model	None about the flow	Targeted: route, split, evaluate	All of them
Number of model calls	Known	Bounded	Variable
Testing	Step by step	Step by step, plus the decision points	Whole scenarios
Good fit	Stable, repetitive process	Stable process with varied cases	Open-ended problems

Anthropic's guide Building effective agents uses agentic systems as the umbrella term and separates workflows, where models and tools are orchestrated through predefined code paths, from agents, where the model directs its own process. The patterns below use the names from that guide.

The five patterns to know

1. Prompt chaining

The task is split into consecutive steps, each model call working on the output of the previous one, with optional checks in between. Example: extract the clauses of a contract, compare them with your terms, then write the summary. It is the least agentic pattern and often the best value; see our article on prompt chaining.

2. Routing

The model classifies the incoming request and sends it to the right handling. Example: a customer email goes to the "order status", "return" or "complaint" branch, each with its own prompt and rules. Routing also lets you send simple requests to a cheaper model.

3. Parallelisation

Several calls run at the same time and their results are combined. Two variants: splitting a task into independent parts (analysing each chapter of a report separately), or running the same question several times to compare answers, for example in a compliance check where any doubt should trigger a human review.

4. Orchestrator and workers

A central model analyses the request, decides which subtasks are needed, hands them to other calls and assembles the result. The difference from parallelisation: the subtasks are not planned in advance, they depend on the request. Example: to answer a tender, the orchestrator identifies the required sections and starts one search per section in your document base.

5. Evaluator-optimiser

One call produces a result, a second evaluates it against explicit criteria and returns feedback, and the loop repeats until the result is accepted or a limit is reached. Example: a customer reply is checked against the tone guide and the commercial policy before an employee sees it. This pattern only makes sense when the criteria are clear.

Pattern	Decision left to the model	Main risk
Chaining	None or almost none	An error carried from step to step
Routing	The choice of branch	Wrong routing
Parallelisation	None, except in voting	Multiplied cost
Orchestrator	How to split into subtasks	Unnecessary or forgotten subtasks
Evaluator-optimiser	Accept or redo	Endless loop, unpredictable cost

Example: handling requests for quotes

A small services company receives quote requests by email every week, of very uneven quality. Here is the process as an agentic workflow:

1. Trigger. An email arrives at the dedicated address.
2. Routing (model decision). The model classifies: complete quote request, incomplete request, other topic. "Other" goes to a person and the workflow stops.
3. Extraction. For a quote request, the model extracts the need, volumes, deadlines and any budget into a fixed format.
4. Check by code. If a mandatory field is missing, the request switches to "incomplete": the model drafts a request for details, which a person approves.
5. Orchestration (model decision). The model determines which services from the catalogue are concerned and starts, for each, a search for prices and terms in the internal documents. Limit: five services at most, read-only.
6. Drafting. The model writes the quote from the extracted and retrieved data only.
7. Evaluation (model decision). A second call checks the quote against a checklist: does every price come from the catalogue, are the terms quoted, is the total consistent? Two rounds at most.
8. Human approval. A salesperson reviews, corrects and sends. Nothing leaves without them.

The path is fixed: you always know at which step a case is. The model decides in only three places, and each is bounded by a limit or followed by a check. That is what makes the process testable: you can verify the routing on fifty real emails without running the rest.

When to choose an agentic workflow

It fits when:

• the process is stable, but requests vary too much to cover everything with rules;
• the number of branches would become unmanageable in a plain workflow;
• quality improves clearly with a review against explicit criteria;
• you must be able to explain what happened in each case.

Stay with a simple workflow if a fixed sequence already gives reliable results. Move to an agent only if the steps really cannot be known in advance, as in open research or troubleshooting. Anthropic's guide gives the same advice: find the simplest solution possible and add complexity only when it improves outcomes. The detailed comparison is in AI agent vs. AI workflow, and which business tasks to automate with AI helps you choose where to start.

What it costs

Every decision handed to the model is one more call, and loops multiply calls. In the example, one quote takes between six and about a dozen calls depending on the number of services and evaluation rounds, against three or four for a workflow without orchestration or evaluation. Two rules keep the bill down:

• Bound everything that can repeat: number of subtasks, number of rounds, budget per case.
• Match the model to the step: a small model for routing and extraction, a more capable one for orchestration and drafting.

Estimate both variants with the LLM API cost calculator before you build: multiply the tokens per call by the number of calls in the average case, then in the worst case.

Guardrails

• A fixed format between steps. JSON or a table, never free text.
• Closed decisions. Routing picks from a list; it does not invent categories. Always include an "I don't know" exit to a person.
• Limits wherever there is a loop.
• Minimal permissions. Read-only until writing is truly necessary.
• Human approval before any irreversible action: sending to a customer, paying, changing data. See human-in-the-loop AI.
• A log per case: the decision taken at each point, with input and output.
• A test set per decision point, rerun on every prompt or model change.

If a workflow touches decisions about people, such as applicants or credit, check the regulatory side first; see AI agents and the EU AI Act.

Common mistakes

• Adding agentic steps everywhere. Every delegated decision is a source of variation. Three well-chosen decision points beat ten.
• Forgetting the exit branch. Without "other" or "incomplete", routing forces unusual cases into a branch that does not fit.
• Leaving the evaluator without criteria. "Check that it is good" produces nothing useful; a checklist does.
• Testing only the final result. You lose the main advantage: knowing which decision was wrong.
• Confusing a demo with production. A workflow that works on five examples has not met your difficult cases yet.

Checklist before you build

• The process is drawn step by step, without AI, as it runs today.
• The points where a model decision adds value are identified and kept to the necessary.
• Every decision has a closed list of options and an exit to a person.
• Every loop has a limit on rounds and budget.
• Irreversible actions go through human approval.
• A set of real cases exists for each decision point.
• Cost is estimated for the average and the worst case.
• An owner reads the logs during the first weeks.

Sources

• Anthropic: Building effective agents (definitions of workflows and agents, five workflow patterns)
• European Commission, AI Act Service Desk: How are AI agents addressed within the AI Act?