Typical Workflow

1. Choose inputs

Select a client/backend and define the task context you need to solve.

2. Instantiate core objects

Create tools, prompt templates, and either DirectLLMCall or MultiStepAgent.

3. Execute or inspect

Run one or more calls and inspect model output, tool calls, and step metadata.

4. Capture artifacts

Persist execution outputs and traces for reproducible comparison.

5. Connect to the next library

Bind runs into study definitions in design-research-experiments and export records that can be interpreted with design-research-analysis.

Choosing The Right Entry Point

Need	Recommended entry point
Plain text generation	DirectLLMCall
Iterative reasoning without tools	MultiStepAgent(mode="direct")
Iterative tool use	MultiStepAgent(mode="json")
Code-action loops	MultiStepAgent(mode="code")
Higher-order orchestration	Patterns

Use DirectLLMCall when you want the shortest execution path and minimal control overhead. Move to MultiStepAgent when interpretability and control matter more than absolute latency. Move to patterns when you need coordinated multi-role behavior and reusable orchestration logic.