AI Workflows with Prompt Flows
Explore how to build predictable, auditable AI pipelines using Amazon Bedrock Prompt Flows. Understand the nine node types, data mappings, batch processing patterns, deployment best practices, and monitoring using CloudWatch and X-Ray to ensure reliable production workflows.
Many production AI systems do not need models that improvise. They need a pipeline that executes the same sequence of steps every time, with full visibility into what happened at each stage. While Bedrock Agents excel at dynamic reasoning, in which the model decides its next action at runtime, a large class of workflows, such as document processing, content moderation, and compliance checks, demand predictable, auditable execution paths defined entirely by the developer.
Amazon Bedrock Prompt Flows addresses this need directly. It provides a visual, node-based interface for constructing deterministic multi-step AI pipelines where each node performs a specific operation and its output feeds directly into the next. Think of it as drawing a flowchart that actually runs. The developer defines the entire execution graph at design time, and the flow follows that graph on every invocation, with optional conditional branching for routing logic. This lesson covers the full set of node types, data mapping between nodes, batch processing patterns using iterator and collector nodes, versioning and aliases for production stability, and monitoring with CloudWatch and X-Ray.
The following table provides a quick-reference framework for deciding between Agents and Prompt Flows:
Amazon Bedrock Agents vs. Bedrock Flows Comparison
Criteria | Bedrock Agents | Bedrock Flows |
Execution Model | Dynamic ReAct loop where the model plans and acts iteratively | Deterministic developer-defined directed acyclic graph (DAG) |
Decision Authority | Foundation model decides next action at runtime | Developer defines every step and branch at design time |
Use Case Fit | Ambiguous multi-tool reasoning with unpredictable user intent | Well-defined multi-step pipelines with repeatable logic |
Auditability | Trace-based reconstruction after execution | Fixed execution path fully auditable before runtime |
Branching | Model-driven based on reasoning output | Explicit condition nodes with developer-defined logical expressions |
Complexity Management | Agent handles orchestration autonomously | Developer designs and maintains the DAG explicitly |
With this distinction clear, the next step is understanding the individual building blocks that compose a prompt flow.
Flow node types
Every prompt flow is a
The following node types are available in prompt flows.
Input node: This is the entry point that receives the flow’s initial payload, such as a document, a user query, or a list of items to process.
Output node: This is the terminal node that returns the flow’s final result to the calling application.
Prompt node: This invokes a foundation model using a prompt ...