HumanLayer provides built-in state preservation to help maintain context across agent interactions. This is particularly useful when building agents that need to maintain conversation history or workflow state across multiple human interactions.

For a runnable example, see the FastAPI Email Example.

​
Overview

State can be preserved in both function calls and human contacts through the state field:

# Store state when creating a function call
function_call = await hl.create_function_call(
    spec=FunctionCallSpec(
        fn="send_email",
        kwargs={"to": "user@example.com"},
        state={
            "conversation_history": previous_messages,
            "workflow_context": current_context
        }
    )
)

# Store state when creating a human contact
contact = await hl.create_human_contact(
    spec=HumanContactSpec(
        msg="Do you approve this draft?",
        state={
            "draft_version": 1,
            "previous_feedback": feedback_history
        }
    )
)

The state object will be preserved and returned in webhooks, allowing your application to restore context when handling responses.

​
Example: Email Thread Management

Here’s a practical example of using state to manage an email conversation thread:

class Thread:
    initial_email: EmailPayload
    events: list[Event]  # Track all events in the conversation

    def to_state(self) -> dict:
        """Convert thread to a state dict"""
        return self.model_dump(mode="json")

    @classmethod
    def from_state(cls, state: dict) -> "Thread":
        """Restore thread from preserved state"""
        return cls.model_validate(state)

async def handle_response(human_response: HumanContact):
    # Restore thread state from the response
    if human_response.spec.state is not None:
        thread = Thread.model_validate(human_response.spec.state)
    else:
        raise ValueError("state is required")

    # Update thread with new response
    thread.events.append(
        Event(type=EventType.HUMAN_RESPONSE,
              data={"response": human_response.status.response})
    )

    # Continue the conversation with preserved context
    await handle_continued_thread(thread)

Here’s how to handle webhook responses with FastAPI background tasks:

@app.post("/webhook/human-response")
async def handle_webhook(
    response: HumanContact,
    background_tasks: BackgroundTasks
) -> Dict[str, str]:
    """Handle webhook responses from HumanLayer"""
    if response.spec.state is None:
        return {"status": "error", "message": "missing state"}

    # Process response in background to avoid webhook timeout
    background_tasks.add_task(handle_response, response)
    return {"status": "ok"}

Important: Version your state objects carefully! If you change your Pydantic models, responses from old webhooks may fail to deserialize. Consider:

  • Adding version fields to state objects
  • Supporting multiple versions of model schemas
  • Using migration functions for old state formats
  • Keeping old model versions around until all pending webhooks are processed

Example of versioned state:

class ThreadStateV1(BaseModel):
    version: Literal[1] = 1
    initial_email: EmailPayload
    events: list[Event]

class ThreadStateV2(BaseModel):
    version: Literal[2] = 2
    initial_email: EmailPayload
    events: list[Event]
    metadata: dict  # New in V2

def migrate_v1_to_v2(v1: ThreadStateV1) -> ThreadStateV2:
    """Migrate old state format to new version"""
    return ThreadStateV2(
        initial_email=v1.initial_email,
        events=v1.events,
        metadata={},  # Set defaults for new fields
    )

def load_thread_state(state: dict) -> Thread:
    """Load thread state with version handling"""
    version = state.get("version", 1)
    if version == 1:
        v1 = ThreadStateV1.model_validate(state)
        return migrate_v1_to_v2(v1)
    elif version == 2:
        return ThreadStateV2.model_validate(state)
    else:
        raise ValueError(f"Unknown state version: {version}")

​
Best Practices

  1. Serializable State: Ensure your state object can be serialized to JSON
  2. Minimal State: Store only what’s necessary to restore context
  3. Type Safety: Use Pydantic models to ensure type safety of state objects
  4. Validation: Always validate state when restoring from webhooks
  5. Error Handling: Have fallbacks for missing or invalid state

​
State vs Run IDs

While run IDs group related operations together, state preservation allows for richer context:

  • Run IDs: Group related operations, useful for audit trails
  • State: Preserve detailed context needed for conversation continuity

Use both in combination for robust agent workflows:

hl = HumanLayer(
    run_id="email-campaign-assistant",  # Group operations
)

await hl.create_human_contact(
    spec=HumanContactSpec(
        msg="Review this campaign draft?",
        state=thread.to_state(),  # Preserve conversation context
    )
)
Response WebhooksIntroduction