Source: github.com/pydantic/pydantic-ai/releases/tag/v2.55.0
The changelog lists Python 3.11 as the new minimum for every package and says Python 3.10 installs resolve to version 2.54.0 and earlier. Several changes are flagged as compatibility notes, including recording tool requests from ExaSearch, YouSearch and LocalStack under durable execution, and adding a logfire_ prefix to LogfireMCP tool names.
New features include Conversation, which stores a conversation between runs and is accepted as conversation= by every entry point, and PostgresStepStore and PostgresMediaStore for hosted Postgres storage of messages and media. A unified cache setting and a Caching capability add cross-provider prompt caching, and claude-haiku-5-5 is now supported. Realtime work includes a hang_up() method for ending WebRTC calls from the server on gpt-realtime and GPT-Live.
Bug fixes cover FallbackModel attempts, which are now recorded with their model, timing, error and usage, along with retry handling for Anthropic requests and MCP tool schemas. The complete list of pull requests and contributors is in the release notes on GitHub.
Why it matters if you build agents
Agents running on Python 3.10 will stay on version 2.54.0 or earlier until their interpreter is upgraded. Teams that want conversation persistence in Postgres or prompt caching across providers now have built-in options to evaluate. Rejected responses from a FallbackModel are now counted in run usage, so cost and error tracking may shift after upgrading.
Get stories like this every weekday
One short email with the day's agent tooling news.