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.

DirectLLMCall

Managed multi-step agent

You need iterative reasoning, tool use, or code-action loops but still want a packaged agent entry point.

MultiStepAgent

Reusable workflow graph

You want to author topology directly from typed steps and reuse that graph across runs.

Workflow

Prebuilt orchestration strategy

A standard coordination form already matches the study design you have in mind.

design_research_agents.patterns

Runnable exemplar

You want a copyable reference implementation, expected outputs, and dependency notes.

Examples Guide

Quick Chooser#

  • Start with DirectLLMCall when latency and a small surface matter more than orchestration control.

  • Start with MultiStepAgent when you want managed iterative behavior without authoring a custom graph.

  • Start with Workflow when 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:

  • SeededRandomBaselineAgent is the lightweight control-condition participant.

  • PromptWorkflowAgent wraps a prompt-mode Workflow when experiment code should own the problem object and run metadata.