Pydantic AI v2.64.0

Summary

Pydantic AI 2.64.0 was released on 2026-10-30. Its main additions are a ModelFallback capability with per-attempt request hooks, a BoxdSandbox for running coding and shell capabilities in a remote machine, and security scanning of memory and skills. It also has several compatibility changes worth checking before upgrading.

Key points

• Model fallback: the ModelFallback capability (also buildable from agent specs) moves to another model on failure, with a per-attempt prepare_model_request hook, RetryModelRequest, and sticky=True to keep a conversation on the model that answered. A new first_event_timeout setting falls back when a stream never starts.
• BoxdSandbox: runs Coder, Shell, FileSystem, and other harness capabilities inside a boxd machine instead of the local host.
• ArtifactScanner: scans Memory and Skills content for security issues.
• Prompt caching: @agent.system_prompt(dynamic=True, on_change='append') appends changes instead of rewriting the cached prompt.
• Harness: Reminder(scope='conversation') keeps reminders in history; RunContext.load_capability loads capabilities on demand; ModelResponse.model_id records which model answered.
• Evals: plain functions work as evaluators; experiment spans record CI provenance (OpenTelemetry vcs.* and cicd.* attributes).
• Compatibility: json-repair is now optional (needed by Coder and RepairToolArguments); Coder tools default to non-strict schemas (Coder(strict=True) to opt back in); stored response and container IDs are only sent to the endpoint that created them.
• Upgrade with pip install -U "pydantic-ai==2.64.0".

Why it matters

Built-in fallback with a first-token timeout covers a common production failure (a provider that accepts the request but never streams), and scanning memory and skills targets a growing prompt-injection path for long-lived agents.