Source: github.com/pydantic/pydantic-ai/releases/tag/v2.63.0
The v2.63.0 release changes several defaults. CodeMode now uses dynamic_catalog=True, so a revealed tool is appended instead of rewriting the cached prompt. The Harness LLMReminder now has a default interval of 10 instead of generating on every request.
New features include a task_timeout option on Dataset.evaluate to bound each task attempt, support for ag-ui-protocol 1.0, and OpenAI gpt-6.1-terra model support. The release also adds an IPython Code Mode alongside the Monty-based CodeMode, a DetectLoops capability that interrupts repeated-action loops, and PostgresSpendStore, which lets harness SpendLimits budgets be shared through PostgreSQL.
Bug fixes include requiring mistralai 3.1.0 or later so MistralModel imports on fresh installs. Rate-limit handling for Azure too_many_requests responses is kept on openai 3.32.0 and later. WebSearchTool and WebFetchTool are no longer offered on AsyncAnthropicBedrockMantle, which rejects them.
Why it matters if you build agents
Developers building agents on Pydantic AI can use loop detection and PostgreSQL-backed spend limits without writing their own. Code that depends on the new defaults, such as CodeMode's dynamic catalog or the LLMReminder interval, should be checked because behavior changes without an API change. Teams that pin mistralai or openai should note the new minimum versions.
Get stories like this every weekday
One short email with the day's agent tooling news.