Skip to content

Pydantic AI 2.63 adds loop detection and shared spend budgets

Updated Pydantic AI

Summary

New evaluation timeouts, IPython Code Mode and PostgreSQL-backed spend storage arrive with changed defaults.

Category
Frameworks
Date
Pricing
Not stated in the sources we cited.

Pydantic AI published 2.63.0 on October 28, 2026.

Bound and observe agent runs

DetectLoops can interrupt repeated-action loops. Dataset evaluation gains a timeout for each task attempt, while PostgresSpendStore lets harness budgets share a PostgreSQL-backed store. These controls do not guarantee that every wasteful or unsafe run is detected.

Review changed defaults

CodeMode now defaults to a dynamic tool catalog, appending newly revealed tools instead of rewriting the cached prompt. The release adds an IPython-based Code Mode alongside the existing Monty implementation. Harness reminders default to an interval of ten rather than every request.

UI adapters can manage message history to avoid sending server-held history twice. The release also adds AG-UI protocol 1.0 support and provider compatibility fixes.

Check the complete compatibility notes before upgrading realtime or custom-capability integrations. Framework features and model-provider charges remain separate; no measured savings are claimed.

See the pydantic-ai hub.

More from Pydantic AI

View all

Questions

What is Pydantic AI?

The product site describes it as a model-agnostic Python framework featuring type safety, structured outputs, validation and observability.

Is Pydantic AI open source?

The GitHub repository reports an MIT license.

Which models does it support?

The product site says it supports OpenAI, Anthropic, Google and more.

Filed on Agent Radar. Claims in this note are limited to the sources above.