pydantic-ai v2.67.0 — v2.67.0 (2026-11-10)

Pydantic AI is a Python framework for building LLM agents from the team behind Pydantic. Version 2.67.0, released November 10, 2026, changes by default which instruction files an agent reads from a code repository. The change is in RepoContext, the component that loads files such as AGENTS.md and CLAUDE.md into an agent's context, and it reaches Coder, the framework's bundled coding agent, which uses RepoContext. Teams that keep both files, or that run Coder on durable execution, should check the release's compatibility notes before upgrading.

According to the main pull request, RepoContext now loads one instruction file per directory, AGENTS.md first and CLAUDE.md as the fallback, so a directory that holds both now contributes only AGENTS.md. Instruction files in subdirectories are now picked up by default: when the agent reads or lists something in a directory, that directory's file is added to the conversation, and a companion change extends this to writing and editing files there. A new walk_up option sets how far up the directory tree loading goes. By default RepoContext looks only in the working directory, while Coder now loads files from the repository root down to the working directory. To bound that extra text, nested files are capped at 32,000 characters per conversation by default, and a file that would pass the cap appears as a pointer naming it rather than its contents. The pull request gives the old behavior back with filenames=('CLAUDE.md', 'AGENTS.md') and nested_traversal=False.

The less visible cost falls on durable execution, where a journal records an agent run's completed steps so that, in the documentation's words, a restarted run "replays the completed ones from the journal." Because Coder now makes extra file checks when a run starts, the pull request warns that its Temporal and DBOS runs started before the upgrade "no longer replay," and advises letting in-flight runs finish or versioning the deployment first. A separate change makes SubAgents with no id fail when the agent is built under DBOS, Prefect and the harness durability engines, as it already did under Temporal.

The rest of the release is a long list described only by pull-request titles. It adds support for Anthropic's Claude Fable 5.2 and Claude Mythos 5.2 models, Google's Gemini 4 Carbon and OpenAI's fast and ultrafast service tiers, along with many fixes to durable execution, sandboxes and provider requests.