autogen-quickstart

AutoGen Quickstart

An AutoGen AssistantAgent answers a returning user grounded in a Statewave-backed memory. The integration is two helpers — both wrap AutoGen's system_message, which keeps the adapter framework-version-agnostic.

What this demo shows

Capability	How it's demonstrated
Memory in system_message	build_system_message prepends a Statewave bundle to the agent's static role/instructions
Refreshable per turn	update_system_message re-fetches a fresh bundle before the next turn — useful for multi-turn chats
Turn-by-turn capture	record_turn writes the (user, assistant) pair back as an episode
No AutoGen subclassing	All helpers are duck-typed; works with AssistantAgent, ConversableAgent, or any agent exposing update_system_message

Prerequisites

A running Statewave server at http://localhost:8100 — the docker-compose.yml in the examples root brings up Postgres 16 + pgvector and the API together:

docker compose up -d         # from statewave-examples/

Dependencies (framework pinned in the example only, not in the core SDK). This example targets the classic AutoGen 0.2 API (from autogen import AssistantAgent); the unpinned pyautogen on PyPI now resolves to a 0.10.x package with a different import surface, so pin the 0.2 line:

pip install "statewave>=0.10.0" "pyautogen==0.2.*"
export OPENAI_API_KEY=sk-...

The newer autogen-agentchat (0.4+) split exposes a different API; the same Statewave pattern (build the system_message, record turns) maps over directly.

Run

python autogen_quickstart.py

The demo seeds Alice's profile, compiles it, then asks an AssistantAgent about her plan and contact channel — with Statewave context already woven into the system_message.

The adapter

adapter.py is three small functions, dependency-free:

• build_system_message(client, subject_id, base_prompt, *, task=..., max_tokens=1000) — build a system_message string with a Statewave bundle on top and base_prompt (the agent's role/instructions) below.
• update_system_message(agent, client, subject_id, base_prompt, *, task, max_tokens=1000) — call mid-chat to refresh the agent's system_message with a fresh, task-biased bundle.
• record_turn(client, subject_id, user_msg, assistant_msg, *, agent_name="autogen") — record one user/assistant turn as a Statewave episode.

Smoke test

The adapter's wiring is covered by mock-based tests — no AutoGen install needed:

pytest test_adapter.py

See also

• Statewave Python SDK — the underlying StatewaveClient.
• Getting started — bring a server up in 5 minutes.
• ../langchain-quickstart/ — same pattern, BaseMemory-shaped, for LangChain.
• ../crewai-quickstart/ — same pattern, Task-shaped, for CrewAI.