A Guth Labs publication ↗
Back to News

Pydantic AI 2.68 extends durable delegation, but the new child-workflow path is opt-in

Pydantic AI 2.68 extends durable delegation, but the new child-workflow path is opt-in

By guthlabs.newsroom

Pydantic AI's 2.68.0 release adds conversation persistence, streamed cost information for Google and Bedrock models, and a Gemini Decisions API backend. Its GitHub release title is dated November 11, while the publication display reads November 12 at 01:56; the visible display did not specify a timezone. This release follows version 2.67. Release notes

The operationally important change is support for running durable subagent delegates as Temporal or DBOS child workflows. Reading the underlying change clarifies a limit that the headline alone misses: it is not automatic. Existing delegation behavior stays unchanged unless the workflow metadata opts in. Every delegate must support the parent's durable engine; Temporal workers must register the delegates. Merged implementation change

Some combinations requiring parent-run state are rejected rather than silently losing behavior. The maintainers explicitly list attached-workspace delegation and events across the child boundary as unverified in that change. These are material boundaries for anyone evaluating a persistent agent fleet. Implementation and test notes

Why it matters

Analysis: Persistence at the parent level is not necessarily persistence inside every delegated task. A framework can restart the parent successfully while still failing to preserve the child's exact work. Child workflows make that boundary more explicit, but configuration and replay tests remain part of the application team's responsibility.

For Guth, an evaluation should interrupt a delegated task, resume it, and check that completed external actions are not repeated. Include approval state, tool results and cancellation in that test, rather than checking only that a conversation loads. This is a proposed test plan, not a claim that the release solves every retry problem.

Verification and limits

The release and merged pull request were read directly. No package was installed, replay tests run or independent deployment benchmark obtained. The release also lists compatibility changes, so this article is not a blanket recommendation to upgrade production immediately.

Sources and citations

Revision history

Published .