Where To Start#
design-research-agents has a deliberate abstraction ladder. Start with the
smallest layer that matches your need, then move upward only when the lower
layer becomes constraining.
Abstraction Ladder#
Layer |
Start here when |
Primary surface |
|---|---|---|
Primitive direct call |
You want one prompt in and one structured result out with minimal control overhead. |
|
Managed multi-step agent |
You need iterative reasoning, tool use, or code-action loops but still want a packaged agent entry point. |
|
Reusable workflow graph |
You want to author topology directly from typed steps and reuse that graph across runs. |
|
Prebuilt orchestration strategy |
A standard coordination form already matches the study design you have in mind. |
|
Runnable exemplar |
You want a copyable reference implementation, expected outputs, and dependency notes. |
Quick Chooser#
Start with
DirectLLMCallwhen latency and a small surface matter more than orchestration control.Start with
MultiStepAgentwhen you want managed iterative behavior without authoring a custom graph.Start with
Workflowwhen the graph itself is the thing you are designing and validating.Start with Patterns when you want a reusable plan/execute, debate, routing, RAG, or related pattern.
Start with Examples Guide when you want exemplars to copy from, not a new abstraction to extend.
Specialized Public Entrypoints#
Two public entry points sit beside the main ladder because they target packaged problem studies rather than general prompting:
SeededRandomBaselineAgentis the lightweight control-condition participant.PromptWorkflowAgentwraps a prompt-modeWorkflowwhen experiment code should own the problem object and run metadata.
Recommended First Click#
New to the library entirely: Quickstart
Choosing between agent entry points: Agents
Building your own reusable topology: Workflows
Reusing a higher-order orchestration strategy: Patterns
Looking for copyable end-to-end examples: Examples Guide