What changed
Pydantic AI 2.28.0 pushes realtime voice from demo territory into the main framework surface. The release adds speech-to-speech support with Agent.realtime(), browser WebRTC plus server sideband support, and a new provider path for live voice interactions.
The product point is not just that Pydantic AI can now talk. Its realtime overview docs make clear that a live session uses the same tools, dependencies, history, usage limits, and observability as the rest of a normal Pydantic AI agent. That means a voice agent can use the same app logic a text agent already uses instead of forcing teams into a separate voice stack.
Why it matters
This is a meaningful shift for builders because speech interfaces usually create a second architecture: one pipeline for chat, another for voice, and extra glue for tools, auth, and tracing. Pydantic AI is trying to collapse that split. If your existing agent can already look up account data, call internal tools, or enforce usage limits, the new realtime path lets you reuse those primitives in a spoken experience.
That makes Pydantic AI more interesting for product teams building customer support, internal copilots, or workflow assistants where voice is an interface choice rather than a separate product line. The release is effectively saying that speech-to-speech should be a first-class mode of an agent framework, not a special integration living off to the side.
The adoption caveat
The same release also ships a high-severity fix for the framework's development web chat UI. Pydantic's security advisory GHSA-h4xc-3qfq-jf93 says a website open in a developer's browser could trigger the local chat endpoint and cause tool execution with the local process's privileges. Version 2.28.0 fixes that by requiring Content-Type: application/json before the agent runs.
That pairing matters. Pydantic is not only adding a new realtime surface; it is also tightening the local developer runtime around agent tooling. For teams evaluating voice agents, that is a useful reminder that framework maturity is not only about adding modalities. It is also about making the local and production safety model harder to misuse.
Practical takeaway
The headline is not simply that another framework added voice. The more important change is that Pydantic AI is moving toward one agent runtime that can cover text, tools, realtime audio, and observability with fewer parallel implementations. If that model holds, teams get a faster path from prototype chat agent to production voice workflow without rebuilding the stack around a second runtime.