Source · release

Pydantic AI v2.69.0

Author
Pydantic
Published
16 November 2026
Added
17 November 2026
Original
https://github.com/pydantic/pydantic-ai/releases/tag/v2.69.0

Summary

Pydantic AI 2.69.0 (2026-11-16) adds tool-call budgets, more sandbox backends including Kubernetes and a Windows workspace, peer messaging between agents, and an autonomy option for its Coder agent. It also adds Gemini 4 Hydrogen support and an Anthropic data-residency setting.

Key points

  • Budgets: ToolBudgets answers over-budget or repeated tool calls with a message to the model instead of raising an error; Advisor.max_uses can be a callable so one budget spans several runs.
  • Sub-agents: SubAgent(keep_partial_output=) passes a sub-agent's last text to the parent when it times out or runs out of budget; DynamicWorkflow(call_scheduler=) controls when sub-agent calls start; Peers in pydantic-ai-harness lets agents message each other.
  • Sandboxes: KubernetesSandbox (kubernetes-sigs agent-sandbox pods), WindowsWorkspace (PowerShell), and network or secret options on the E2B, Boxd, Daytona, and OpenSandbox backends.
  • Coder: a new autonomy option and shorter default instructions; benchmark-only instructions moved to an opt-in BenchmarkInstructions.
  • Models and settings: Gemini 4 Hydrogen on GoogleModel, service_tier on GroqModel, anthropic_inference_geo for Anthropic data residency, and openai_safety_identifier. Compaction re-injects pinned context as framed data (pin(source=)).
  • Install or upgrade with pip install -U pydantic-ai.

Why it matters

Budgets that degrade gracefully and partial output from failed sub-agents make multi-agent runs less brittle; Kubernetes sandboxes make it easier to run agent code isolated on existing clusters.