design-research-agents

A flexible, modular framework for researching AI agents in design workflows.

Build and compare agent behaviors, swap LLM backends, and capture traces without rewriting your pipeline. The library favors small, composable pieces so you can test ideas quickly and keep experiments reproducible.

Highlights

• Two core agent entry points: DirectLLMCall and MultiStepAgent.

• MultiStepAgent supports explicit modes: direct, json, and code.

• JSON mode uses structured tool_name/tool_input selection for iterative tool-call loops.

• Model selection policies with local/remote catalogs.

• Tool contracts and schemas for safe, structured I/O.

• Tracing hooks and emitters for debugging and evaluation.

• Runnable examples for deterministic validation and experimentation.

• Workflow-native memory, networked blackboard coordination, and reusable reasoning patterns (tree search and RAG).

Typical workflow

1. Choose an agent type and backend.

2. Define tools, prompts, and policies.

3. Run experiments and capture traces.

4. Compare results and iterate.

Get started

• Quickstart for a fast, end-to-end example.

• Dependencies and Extras for optional dependency profiles and platform constraints.

• Examples Guide for scenario-driven runnable examples and expected observations.

• Workflow Primitive Examples for runnable workflow primitive examples.

• Pattern Examples for runnable orchestration pattern examples.

• LLM Clients to choose local or remote client backends.

• Tools for unified runtime + MCP + script tools.

• Agents to understand agent execution tradeoffs.

• Workflows for workflow builder primitives and composition.

• Patterns for prebuilt workflow implementations.

• API for the guaranteed public API surface.

• CONTRIBUTING.md for contribution workflow and PR expectations.