Ship agents that behave well, consistently.

Parlant contextually matches your most relevant rules to each turn of the conversation, maximizing the language models' groundedness and instruction-following consistency.

# Create a hook for refund-related rules and tools
refunds = await agent.create_observation("discussing refunds")

find_order_id = await agent.create_guideline(
  condition="The refund request lacks an order ID",
  action="Find and confirm the order ID before continuing",
  # Allow the use of these tools when this guideline is active
  tools=[find_order_by_id],
  # Ignore lower-priority guidelines until this one is resolved
  priority=10,
  # This guideline can only activate in the scope of refunds
  dependencies=[refunds],
)

A clean structure for your conversations.

Every conversation session is a timeline of events, not a fragile request-response loop.

Why Parlant (2)Routed graphs fail to deliver

What Parlant does instead

journey = await agent.create_journey(
    title="Wire Transfer",
    conditions=["customer requests a wire transfer"],
)

# Each state names what to do next, not how to route.
t1 = await journey.initial_state.transition_to(
    chat_state="Ask for recipient name and amount",
)
t2 = await t1.target.transition_to(
    condition="received name and amount",
    tool_state=verify_recipient_and_balance,
)

await agent.create_guideline(
    condition="customer requests a refund",
    action="process the refund using their order ID",
    tools=[process_refund],
    composition_mode=p.CompositionMode.STRICT,
    canned_responses=[
        await server.create_canned_response(
            template="Refund processed. Transaction ID: {{transaction_id}}",
        ),
    ],
)

{
  "turn": 4,
  "matched_guidelines": [
    {"id": "require_order_id",   "score": 0.94},
    {"id": "polite_acknowledge", "score": 0.71}
  ],
  "canned_response": "acknowledge_then_ask_order_id",
  "tools_called": ["lookup_order"],
  "latency_ms": 412
}

What the pros say

From the teams running against real customers, at real volume, in domains where being wrong is expensive.

• “

  We tested Parlant extensively. The failure patterns that exist in our production logs, capturing them in Parlant takes minutes.

• “

  Parlant isn't just a framework. It's a high-level software that solves the conversational modeling problem head-on.

• “

  Parlant dramatically reduces the need for prompt engineering and complex flow control. Building becomes like modeling.

• “

  We went live with a fully functional agent in one week. I’m impressed by how consistent Parlant is with its responses.

Frequently asked

• 01What is a harness?+

  A harness performs automatic context engineering on top of a language model. Context engineering is the discipline of getting the right context, no more, no less, into the prompt at the right time. You can't cram all of your instructions, knowledge, and tools into one prompt regardless of the model or context window size. Models don't listen beyond a certain point. You need the right items in the prompt, and the wrong items out. Parlant is a conversational harness: it assembles only the relevant rules, data, and tools for each turn.

• 02How is this different from prompt engineering?+

  Prompt engineering is about getting a model to behave correctly when it already has the right context: optimizing the template, tuning temperature, tweaking reasoning. Context engineering is about what goes into that template in the first place. A well-engineered prompt with the wrong context still produces wrong results. They are complementary: prompt handles how, context handles what and when.

• 03How is Parlant different from LangChain or LangGraph?+

  LangChain and LangGraph are workflow tools, they manage how data flows between steps. Parlant works on a different level: it manages behavior. Instead of building a graph of nodes and edges, you define individual behavioral rules (guidelines) and the relationships between them. The framework figures out which rules apply at each turn. Changing your agent's behavior is a content change (add a guideline), not a structural change (restructure the graph).

• 04What about scenarios I didn't write rules for?+

  The LLM still handles general conversation naturally. Guidelines define behavioral expectations for specific situations, everything else works as you'd expect from an LLM. If you don't need special handling for a scenario, you don't need to define any rules for it.

• 05How do I prevent hallucinations in critical moments?+

  Use canned responses with strict composition mode for control of wording. Since each response can reference fields (from tool results, retrievers, or guidelines), if a required field isn't present in the current context, Parlant automatically disqualifies that response from being selected. The guardrails are structural and deterministic, not just prompt-based.

• 06Won't canned responses make conversations feel rigid?+

  The real rigidity of traditional chatbots comes from tree-based flows, not from controlled wording. Even in strict canned-response mode, the agent still chooses when to use them based on the fluid nature of the interaction, just like a call-center rep does. The flow stays flexible; you get precise wording where it matters.

• 07Does this work with my existing LLM?+

  Parlant is LLM-agnostic, you can use any provider and model. When consistency and reliability matter, some models perform better than others. The officially recommended providers are Emcie, OpenAI (directly or via Azure), and Anthropic (directly or via AWS Bedrock).

Get started

Install. Define your first guideline. Talk to your agent.

$ pip install parlant

End of the manual.