4.2025

We have refined the sidebar layout to better group tools by function and clarify where key capabilities live. Think of this as a smart re-balancing of where things go and a review of section naming.

What’s new, renamed, or moved:

Analyze renamed to Manage

• Agent Analysis → New feature. See below for more details.

Build

• Agent → Renamed from About.
• Model Training → Moved here from Annotations, previously under Conversation Review.
• Variant Management → Moved here from the Configure menu.

Voice

• Agent Voice → Moved here from Configure (combines features previously under Voice and Rules)
• Cache Management → Previously the Audio Management tab.

Configure

• Environments and Project History → New sections, previously under Environment & Versions.

Why it matters

These adjustments make it easier to find what you need, reduce clutter, and prepare the platform for future capabilities.

You can now generate LLM-powered performance metrics in the PolyAI Agent Studio using the new Agent Analysis flow. Define a custom prompt and category set, then run it across a batch of calls to generate insight-rich evaluations and visual summaries.

How it works:

• Create a custom evaluation prompt, with categories and descriptions
• Select a batch of calls using the Agent Analysis feature.
• Automatically assess agent performance across key custom criteria like tone, intent success, or compliance (or any other potential variables)
• View results in a dedicated Agent Analysis UI, with graphs showing how your calls measure up against each evaluation axis

Agent Studio now supports real-time collaboration and edit tracking across your team.

Multi-user refresh view in action

How it works:

• See who else is editing a draft in real time.
• Receive notifications when changes are saved by others.
• Copy unsaved edits before refreshing to avoid overwriting.
• View full change logs using the history panel.

You can now use an LLM to review and rate calls.

What you can do:

• Rate calls for tone, intent success, or compliance
• Use project-specific categories
• Get automatic summaries

You can now filter conversations by workspace on the Conversations page.

Why it helps:

• Focus on the calls your team is responsible for
• Cleaner review experience for large orgs

You can now extract structured address data from free-text input using a built-in LLM utility: extract_address().

What it does:

• Parses the most recent user message for address info
• Returns an Address object with fields like street, postcode, and country
• Raises an ExtractionError if parsing fails

Usage:

try:
    address = conv.utils.extract_address(country="US")
    conv.state.parsed_address = address
except ExtractionError as e:
    conv.state.address_error = str(e)

Address object:

@dataclass
class Address:
    street_number: Optional[str]
    street_name: Optional[str]
    city: Optional[str]
    state: Optional[str]
    postcode: Optional[str]
    country: Optional[str]

Other details:

• Calls an LLM, so may take a few seconds
• You can optionally validate against a list of known addresses
• Some fields may be missing depending on input quality

You can now access earlier conversation turns from within utility functions using conv.history.

What it does:

• Returns a list of events: user inputs, agent replies, function calls, etc.
• Useful for writing context-aware logic

Usage:

def some_function(conv: Conversation):
    history = conv.history
    for event in history:
        print(event.to_dict())
    return {
        "utterance": "Here's a response that depends on previous turns.",
        "content": f"Conversation so far: {str(history)}"
    }

Why it matters:

• Build logic based on what’s already happened in the conversation
• Avoid passing state manually
• Better debugging and flow control

You can now export and import variants between agents or environments.

What you can do:

• Export all variant data to CSV
• Edit things like contact numbers or opening hours in bulk
• Reimport to update the same or a different agent

Why it matters:

• Makes it easy to scale across multiple sites
• No need to rebuild variant logic from scratch
• Keeps your CSVs in sync with other tools

Reviewed conversations now show the variant ID used in each call.

Why it matters:

• See exactly which variant was active
• Easier to debug and compare different versions
• Adds context for QA and bug tracking

The built-in handoff now supports handoff reasons and custom utterances, and you can now use a handoff() method directly inside functions.

What you can do:

• Specify a reason for the handoff (e.g. compliance, escalation, etc.)
• Set a custom utterance to be spoken at handoff
• Use handoff() programmatically inside a function.

Example usage:

    def escalate(conv: Conversation):
        return conv.call_handoff(
            reason="policy_violation",
            utterance="Let me transfer you to a specialist who can help."
        )

You can now configure multiple TTS voices for your agent and distribute them across calls. Use percentage values adding up to 100% to distribute the likelihood of each agent answering a call.

Why it matters:

• A/B test different voices for performance or tone
• Assign different voices to different intents or caller types
• Create more natural, human-like team dynamics

What you can do:

• Select and assign multiple voices
• Preview how your agent sounds across scenarios