What separates a workflow from an agent in Anthropic's guidance?
Why C is the answer
Workflows follow predefined code paths
Workflows orchestrate LLMs and tools through predefined code paths. Agents decide their own next steps from what they find, which is why they cost more and are harder to test.
In detail
- Workflows orchestrate LLMs and tools through predefined code paths.
- Agents decide their next steps from what they find.
- That freedom is why agents cost more and are harder to test.
- Choose by how much the path varies.
Use case
An illustrative scenario, not a real incident.
A team compares an invoice flow with fixed steps and a research assistant that chooses its own searches. The first is a workflow, and the second is an agent, because its path depends on what it finds.
See it in different ways
mindmap
root((Patterns))
Prompt chaining
Fixed steps
Routing
Classify first
Parallelization
Sectioning
Voting
Orchestrator workers
Subtasks found at run time
Evaluator optimizer
Feedback loopflowchart TD
Q{"Who decides what happens next?"}
Q -->|"Your code"| A0["Workflow"]
Q -->|"The model, from what it finds"| A1["Agent"]Key point: In a workflow, your code picks the steps.
Why the other options are wrong
- A. Agents run only as batch jobs
- Agents are not limited to batch jobs.
- B. Workflows never call a model
- Workflows do call models, at defined steps.
- D. Agents run as a single model call, while workflows can use several tools
- An agent usually runs a loop of many calls, not a single one.