README.md

Python examples

The fastest way to understand Agent RT is to run it. These small examples are deliberately incremental: start with one agent call, then add tools, streaming, structured output, runtime limits, and direct registry control without introducing a second framework or hidden abstraction.

If you are evaluating Agent RT, run 01_basic_agent.py first, then 02_tool_calling.py. Together they show the core model → tool → model loop with the same contracts used by the larger runtime.

Examples

• 01_basic_agent.py — minimal agent run with a configured provider.
• 02_tool_calling.py — register a read-only tool and let the model call it.
• 03_streaming.py — consume text deltas with run_streaming().
• 04_structured_output.py — require schema-validated JSON output with one repair attempt.
• 05_run_limits.py — bound turns, tool calls, elapsed time, and total tokens.
• 06_tool_registry.py — use namespaces, direct execution, and enable/disable controls without a model provider.
• 07_vector_db.py — query Chroma, Milvus, Pinecone, Qdrant, or Weaviate and switch backends with environment variables.

Setup

From the python/ directory:

uv sync

Or with pip:

python -m venv .venv
source .venv/bin/activate
python -m pip install -e .

Copy agent-rt.example.yaml to agent-rt.yaml (or ~/.agent-rt/config.yaml) and edit model_list. Secrets are referenced as os.environ/NAME, so export the variables the file names (for example OPENAI_API_KEY).

Run an example from the python/ directory:

uv run python examples/01_basic_agent.py
uv run python examples/02_tool_calling.py
uv run python examples/03_streaming.py
uv run python examples/04_structured_output.py
uv run python examples/05_run_limits.py
uv run python examples/06_tool_registry.py
uv run python examples/07_vector_db.py

The first five examples call a configured model provider. 06_tool_registry.py is fully offline and is useful for checking local setup without API credentials. 07_vector_db.py performs live vector retrieval and requires a configured vector service plus an embedding-capable model provider.

Vector DB example

The vector DB example uses the same application code for every built-in service backend. Add a vector_store_registry entry (backend via custom_llm_provider: chroma, milvus, pinecone, qdrant, or weaviate; plus api_base, collection, api_key: os.environ/..., and an embedding_model that is an embedding-capable model_list entry), then pass the store name and an optional query:

python examples/07_vector_db.py docs-kb "What is Agent RT?"

Change custom_llm_provider plus the endpoint, collection, and credential to move the same application code to another backend; several stores can be declared side by side. Extra non-secret backend settings go under options.

FAISS and pgvector remain custom VectorDBProviderRegistry integrations because their normal execution paths require local/native or database drivers.