Skip to main content

Patterns

These are the core patterns for building agentic systems — how agents coordinate work, reason through problems, use tools, and produce output. They sit between agent creation (how agents are structured) and evaluation (how output quality is measured).

The four categories below move from the outside in: how work flows between components, how agents think, how agents interact with the world, and how output is shaped.


Orchestration

How agents and steps coordinate. These patterns define the flow of work — who does what, in what order, with what control.

PatternDescription
OrchestratorDynamic task delegation to specialized workers
PipelineFixed sequential steps with validation gates
ParallelConcurrent tasks with aggregation
RouterOne-shot classification and dispatch
Evaluator-OptimizerGenerator + critic in a refinement loop
Human-in-the-LoopCheckpoints for human approval

Reasoning & Planning

How agents think. These patterns define the reasoning strategy — single-pass, iterative, planned, or self-correcting.

PatternDescription
Chain of ThoughtStep-by-step reasoning in a single pass
ReActThought-action-observation loop with tool use
Plan-and-ExecuteUpfront planning then step-by-step execution
ReflectionSelf-critique and revision loop

Tool Integration

How agents interact with the external world. These patterns define how agents discover, select, and use tools.

PatternDescription
Tool RouterSelect and invoke the right tool via function calling
RAGRetrieval-augmented generation from knowledge bases
Code ExecutionGenerate and run code in a sandbox
MCPStandardized protocol for dynamic tool discovery

Output

How agents format and validate responses. These patterns define the shape, safety, and delivery of output.

PatternDescription
Structured OutputSchema-constrained typed output
StreamingProgressive token and event delivery
GuardrailsSafety, compliance, and quality validation
CitationSource attribution and grounding