Source · release
Pydantic AI v2.67.0
Summary
Pydantic AI 2.67.0 (2026-11-10) extends durable execution to sub-agents, makes the repository context loader pick up nested instruction files, and adds the newest models from Anthropic, Google, OpenAI, and Moonshot.
Key points
- Durable sub-agents:
SubAgentsacceptsdynamic_agentsandtask_controlunder Temporal and DBOS. With DBOS, tool calls can run as child workflows, in parallel when a workspace is attached.TrajectoryJudgeevaluations are checkpointed when the judge has anid. - Nested instructions:
RepoContextloads one instruction file per directory (with awalk_upoption) and shows nested files when the model edits in that directory, capped bynested_max_chars.Workspace.authority_ofsets per-path authority. - Models: Claude Fable 5.2 and Mythos 5.2 (also on Bedrock and OpenRouter), Gemini 4
Carbon on
GoogleModel,gpt-image-2.6-flareand-sunburst, Kimi K3 on Bedrock; OpenAIfast/ultrafastservice tiers; Anthropicweb_search_20260318andweb_fetch_20260318throughprovider_settings. - Other:
dequeueandset_enqueue_prioritychange queued messages before delivery;KeenableSearchadds keyless web research. - Upgrade:
pip install -U pydantic-ai.
Why it matters
Nested instruction loading mirrors how coding agents read per-folder CLAUDE.md or
AGENTS.md files, and durable sub-agents let long multi-agent runs survive restarts.