Source · release
Pydantic AI v2.71.0: harness capabilities, permission rules, goals
Summary
Pydantic AI 2.71.0 adds a set of ready-made harness capabilities for long-running
agents, rule-based permissions for tools and shell commands, a Goal that keeps
unattended runs going until a check passes, and a ToolBridge that keeps large tool
sets cache-friendly. The release title is dated 2026-11-18; the feed entry appeared on
2026-11-19.
Key points
- Harness capabilities:
ReportToolErrors,WrapUpAtLimits,ReportRemainingBudget,RequireVerification,ReportStaleFiles,CurrentTime,LabelCrossModelHistoryandDenyUnanswered, added to an agent instead of being hand-written into prompts. PermissionRule: allow, ask about, or deny tool calls and shell commands by rule, similar to coding-agent permission settings.Goal: keeps an unattended run going until a caller-defined completion check passes.ToolBridge: exposestool_search,tool_describeandtool_callin place of the full tool list, so the prompt cache survives large or changing tool sets.Shell(secret_env=...)masks injected secrets in tool output;ModelRouterroutes runs between models;GoogleModelsupportsgemini-4-neon.- Breaking:
PeerNetworkis renamedPeerConnection. StackOne'sapi_keyis deprecated in favour ofauth. - Upgrade:
pip install -U pydantic-ai.
Why it matters
These are the building blocks of a coding-agent-style harness (permissions, goals, budget awareness, verification) offered as library pieces, so custom agents can get the same safeguards without reimplementing them.