bigonotes
AIWorkflows

What a Workflow Is

A workflow is a pipeline your code controls, with the model handling some steps. How it differs from tool calling and from an agent.

2 min readLast updated

A workflow is a system where your code decides the steps and the model handles some of them. You write the pipeline: step 1, then step 2, then step 3. The model never chooses what to do next. It only carries out the step your code hands it.

Your code decides#

That one rule is what makes a workflow useful.

  • Predictable. You know exactly what will run, and in what order, every time.
  • Testable. Each step has a fixed input and output, so you can test and evaluate it on its own.
  • Cheap. Only the steps you planned call the model. Nothing runs extra calls on its own.

Workflow vs tool calling#

The two are often confused, but they sit at different levels.

Tool calling Workflow
What it is A mechanism An architecture
Scope One round trip: the model asks, your code runs the function, the result goes back The whole sequence of steps your code runs
Uses the other? No Some steps may use tool calling. Others are plain prompts with no tools.

So the defining feature of a workflow is not tool calling. It is that your code controls the flow.

Workflow vs agent#

In a workflow, your code fixes the steps before the request arrives; in an agent, the model picks each next step at runtime

In an agent, the model decides the next step at runtime. In a workflow, your code decided it before the request arrived.

The five patterns#

Most workflows are built from five patterns. Each has its own note.

Pattern Use it for
Prompt chaining Fixed stages, one after another
Routing Classify the input, then send it to a specialized prompt or model
Parallelization Independent tasks at the same time, or voting
Orchestrator-workers A model splits the task into sub-tasks at runtime
Evaluator-optimizer Draft, critique and revise until quality is met

Short line#

A workflow is a pipeline your code controls, where the model handles some steps but never picks the next one. That makes it predictable, testable and cheap. Tool calling is a mechanism a step may use. In an agent, the model picks the next step at runtime.

Key takeaways#

  • In a workflow, your code decides the steps. The model only carries out the step it is given.
  • Workflows are predictable, testable and cheap.
  • Tool calling is one round trip. A workflow is the architecture of steps around it.
  • In an agent the model decides the next step at runtime. In a workflow your code decided it beforehand.
  • The five patterns are prompt chaining, routing, parallelization, orchestrator-workers and evaluator-optimizer.
Revision notes · 1 min

Core idea#

A workflow is a system where your code decides the steps and the model handles some of them. The model never picks the next step.

Key points#

  • Predictable: the same steps run in the same order every time.
  • Testable: each step has a fixed input and output.
  • Cheap: only the planned steps call the model.

Compare#

Decides the next step When
Workflow Your code Before the request arrives
Agent The model At runtime

Tool calling is a mechanism: one round trip. A workflow is the architecture. Some steps use tool calling, others are plain prompts.

Five patterns#

Pattern Use it for
Prompt chaining Fixed stages in order
Routing Classify, then dispatch to a specialized prompt or model
Parallelization Independent tasks at once, or voting
Orchestrator-workers A model splits the task at runtime
Evaluator-optimizer Draft, critique, revise

Pitfalls#

  • Do not define a workflow by tool calling. Define it by who controls the flow.

Short line#

A workflow is a pipeline your code controls. In an agent, the model picks the next step.

Found a mistake? Email ask@bigonotes.com