Important
Current monthly release: The April Alignment
Due: April 1, 2026
Tracks: March 2026 work
design-research-agents is a modular framework for prototyping and researching engineering design AI agents.
It features shared runtime contracts, workflow orchestration, and pluggable LLM backends for quick iteration.
This library centers on a small set of composable pieces you can run, inspect, and test:
- Two primary entry points:
DirectLLMCallandMultiStepAgent(direct,json, andcodemodes) - A tool runtime built around
Toolbox, with callable, script, and MCP-backed tool configs - Prebuilt orchestration patterns for plan/execute, debate, propose/critic, routing, beam search, RAG, blackboard, and conversations
- Hosted and local LLM clients, plus
ModelSelectorfor backend-selection policies - Tracing, structured
ExecutionResultoutputs, and runnable examples aimed at repeatable experiments - A workflow runtime with explicit step primitives for model calls, tool calls, delegation, loops, and memory
from design_research_agents import LlamaCppServerLLMClient, MultiStepAgent
with LlamaCppServerLLMClient() as llm_client:
agent = MultiStepAgent(mode="direct", llm_client=llm_client, max_steps=3)
result = agent.run(
prompt="Suggest two design goals for a field-repairable drone battery latch.",
)
print(result.final_output)Requires Python 3.12+.
Reproducible release installs target Python 3.12 (see .python-version).
python -m venv .venv
source .venv/bin/activate
make dev
make test
PYTHONPATH=src python3 examples/patterns/plan_execute.pyFor frozen installs, optional extras, and release maintenance, see Dependencies and Extras.
Start with examples/README.md for runnable examples grouped by agents, clients, workflows, patterns, model selection, and tools.
See the documentation site for the full guide set, including quickstart, backend setup, workflows, patterns, and API reference.
Build docs locally with make docs.
Contribution guidelines now live in CONTRIBUTING.md.