Parlant

Parlant is an open-source AI agent framework built around one idea: instead of hoping an LLM follows your 47-rule system prompt, you teach it principles and the framework ensures it. You write behavioral guidelines as condition/action pairs (for example, "customer asks about refunds" triggers "check order status first"), and Parlant's guideline matching engine contextually matches the relevant guidelines on every turn, prepares a fully-aligned response plan, and explains why each guideline was matched.

Around that core it layers customer journeys (scripted multi-step interaction paths), canned responses that eliminate hallucinations on regulated answers, domain glossaries, and tool attachments bound to specific interaction events. The README positions it directly against LangGraph and DSPy as the compliance-first alternative, and at 18,000+ stars it has real traction with teams building support and sales agents where a wrong answer costs money. The honest trade-off is a new mental model to learn: guideline authoring is a discipline, and a badly-written guideline set is just a system prompt with extra steps.

1. Install the package: pip install parlant (Python 3.10+).
2. Export your model provider key (e.g. OPENAI_API_KEY).
3. Start the Parlant server and create an agent via the Python SDK.
4. Teach it principles with await agent.create_guideline(condition=..., action=..., tools=[...]).
5. Attach tools, define journeys for key flows, and connect your chat frontend to the server.