A Guth Labs publication ↗
Back to News

Pydantic AI 2.67 adds durable dynamic agents and records guardrail decisions for replay

Pydantic AI released version 2.67.0 on November 10, extending its agent framework’s support for recoverable workflows while changing several compatibility requirements. Among the changes are durable execution for dynamically created subagents, persisted guardrail verdicts and revised repository-instruction loading. Release notes

The dynamic-agent change addresses a specific recovery problem: created agents previously existed only in process memory. A replacement worker or replay could lose that roster. The implementation now reconstructs it from workflow state and passes a description of the created agent into recorded delegation operations. The linked change reports recovery tests for Temporal and DBOS. It still excludes dynamic agents on engines that record each function tool separately, including Prefect, Lambda, Azure and Render. It also requires an identifier for durable SubAgents configurations that previously allowed id=None; developers can retain the default identifier or supply a distinct one. Dynamic-agent implementation

Guardrail decisions now receive similar treatment. Previously, replay could invoke an input, output or tool guard again, potentially repeating a paid classifier call or producing a different verdict. The new implementation records those results so replay can reuse them. Durable agents must give each guardrail an explicit, distinct identifier. The pull request describes serialization limits, including reduced context available to guards running on Temporal workers and limitations around custom exception types. Guardrail implementation

The upgrade also changes how repository guidance enters an agent’s context. Release notes describe one instruction file per directory, nested files loaded by default and a default 32,000-character cap on inlined nested instructions per conversation. Provider-specific native-tool fields move to provider_settings. These are migration details rather than merely extra model choices, and they merit review against an application’s existing configuration. Compatibility notes

The operational benefit to evaluate is continuity after interruption, not a promise that an agent will complete every task correctly. Teams adopting the release should test recovery with their actual workflow engine, confirm guardrail identifiers and serialized data, and check which repository instructions reach the model. The upstream tests establish the maintainers’ intended behavior; this report did not execute the release or independently validate production reliability.

Sources and citations

Revision history

Published .