How langwatch/create-agent-app works
Project type and purpose
• This is a mono‑repo of multiple “Customer Support Agent” examples, each built with a different LLM‑agent framework (DSPy, Google ADK, InspectAI, LangGraph functional & high‑level, Pydantic AI, smolagents).
• Its purpose is to provide side‑by‑side templates showing how to implement the same agent logic in each framework, sharing common mocks and knowledge‑base data.
File structure
• Top‑level directories:
• create_agent_app/common – shared code (mocked APIs, knowledge base)
• dspy_example, google_adk_example, inspect_ai_example, langgraph_functional_api_example, langgraph_highlevel_api_example, pydantic_ai_example, smolagents_example – each contains one agent implementation, its pyproject.toml, tests, and lockfile.
• Root files: .env.example, Makefile, uv.lock, pyproject.toml – define environment, build/test scripts, dependency management.
• Tests live under each example/tests and drive the agent workflow.
• The common folder holds static knowledge_base markdown files and a mocked_apis.py to simulate external systems.
README for additional insights
• Describes the process: clone → cd into an example → uv sync → pytest.
• Emphasizes that examples should remain self‑contained and nearly identical in behavior, differing only in framework calls.
• Notes all use the same Gemini model endpoint and Scenario tests.
a. Main components:
• Example Agents (one container per framework)
• Shared Common Library (mocked‑apis + knowledge_base)
• External LLM API (Google Gemini)
• Test Harness (pytest driven by uv)
• CI/Build Orchestrator (Makefile / uv commands)
b. Relationships & interactions:
1. Test Harness invokes a given Example Agent.
2. Example Agent imports shared logic from Common Library.
3. Agent logic calls the external LLM API with prompts assembled from knowledge_base.
4. External LLM returns responses, Agent processes and returns to Test Harness.
5. CI/Build Orchestrator orchestrates dependency install (uv sync), test execution, and linting.
c. Architectural patterns & design principles:
• Modular per‑framework containers – supports plug‑and‑play and side‑by‑side comparison.
• Layered separation:
1. Presentation (Test Harness)
2. Application (Agent code in each example)
3. Shared Service (Common library)
4. External Systems (LLM, real APIs vs mocks)
• Single Responsibility: each example only handles framework‑specific orchestration, while common code centralizes mocks and static data.
d. Key technologies:
• Python packaging: pyproject.toml, uv.lock
• Test runner: pytest (invoked via uv)
• LLM endpoint: Gemini‑2.5 flash preview
• Framework SDKs: dsp, google adk, inspect_ai, langgraph, pydantic_ai, smolagents
Elements
• From Test Harness → Example Agent
• From Example Agent → Common Library
• From Example Agent (or Common) → External LLM API
• From CI Orchestrator → Test Harness & Install step