Framework Update | Pydantic AI v2.40.0

TL;DR Pydantic AI v2.40.0 highlights: (1) @agent.on_event decorator gives Agents native event listening; (2) realtime voice sessions now handle barge-in natively; (3) RealtimeSession.enqueue() lets external code inject out-of-band prompts. No breaking changes.

#ai-agent #framework #daily #pydantic-ai

Table of Contents

Version Info
Why This Version Matters
Key Changes
Breaking Changes
Migration Guide
Cross-Framework Observations
Takeaway
References

Series: AI Framework Changelog (16 / 16)

Framework Update | Agno 3.0.6

Version Info

Item	Value
Framework	Pydantic AI
Version	v2.40.0
Previous	v2.39.0
Released	2026-09-05
Release Notes	GitHub Release
GitHub	pydantic/pydantic-ai
Stars	19,742

Why This Version Matters

Pydantic AI takes a significant step forward for voice Agent UX. Previously, handling user barge-in (interrupting the AI mid-speech) required manual plumbing — detecting the interruption, tracking played bytes, and signaling the model to stop. v2.40.0 bakes this into the session layer with a single handle_barge_in=True flag. Meanwhile, the new @agent.on_event decorator gives Agents native event listening for cross-cutting concerns like billing, observability, and live UI updates without monkey-patching internals.

Key Changes

@agent.on_event event listener decorator: Register event callbacks directly on an Agent covering tool calls, token streaming, and completion events → billing and observability logic hooks directly onto the Agent without intercepting internal mechanisms
Realtime voice barge-in handling: handle_barge_in=True + interrupt(played_bytes=...) + played_audio_bytes property → automatic interruption handling when users speak over the AI, with played-progress tracking built in
RealtimeSession.enqueue(): Lets external code driving the session inject out-of-band prompts → enables backend-initiated messages during voice conversations (e.g., a support system pushing a promo code hint)
respond= parameter: RealtimeSession.send(respond=True/False) controls whether a text message solicits a model reply → precise control over multi-turn voice conversation pacing
provider_factory for infer_realtime_model: Custom provider factory function → more flexible realtime model initialization
prices.update_in_background(): Background model pricing data refresh → long-running Agents track latest pricing without restarts

Breaking Changes

No breaking changes in this version.

Migration Guide

Direct upgrade, no code changes required.

pip install --upgrade pydantic-ai==2.40.0

Example using the new event listener:

from pydantic_ai import Agent

agent = Agent("openai:gpt-4o")

@agent.on_event
async def log_events(event):
    print(f"Agent event: {event.type}")

result = await agent.run("Hello")

Cross-Framework Observations

Pydantic AI continues to widen its lead in realtime voice capabilities. Among the 12 Agent frameworks we track, only Pydantic AI offers framework-level native barge-in and audio stream control — other frameworks (LangGraph, CrewAI, Mastra) still require manual WebSocket and audio processing for voice Agents. The @agent.on_event pattern brings the callback mechanism long available in the LangChain ecosystem into a more Pythonic form.

Takeaway

I previously assumed voice Agent barge-in handling was simply "detect user speech and stop." Seeing Pydantic AI's implementation reveals the need to track played byte counts — the model needs to know where the user interrupted to maintain conversation coherence. This is an audio-streaming-specific state management problem with no text-chat equivalent.