Framework Update: Pydantic AI v2.56.0

🌏 繁體中文版

Version Info

Why This Release Matters

The previous release added the Conversation object and Postgres step/media stores; this one fills in the execution environment: harness capabilities can be locked inside a local microVM, and durable execution gains Postgres and Render options. Teams that want Pydantic AI agents in production without committing to a specific workflow engine now have more choices.

Key Changes

• SmolSandbox: a new workspace that runs harness capabilities inside a local smolvm microVM → an isolation option for letting agents run commands without touching the host, and without a cloud service
• PGTaskDurability: durable execution built on pgtask → teams already on Postgres can skip standing up an engine like Temporal
• Render Workflows integration: Render Workflows added as a durable execution engine → agents deployed on Render can use the platform's workflows directly
• Layered Skills: Skills can stack project and personal libraries, name each skill directory, and render a skill as a host command → shared team skills and personal skills can be managed separately
• Models and runtime: support for Gemini 4 Argon (gemini-4-argon) with long decode past the per-request output cap; the Google providers' default timeout is now 20 minutes; Python 3.15 support; ZaiProvider accepts base_url
• Harness Coder capability: WarnNearLimits removed
• Notable fixes: MCPToolset(http_client=...) now applies auth to requests and no longer closes the user-supplied client; a 200 response with a text/plain body from OpenAI Responses, Anthropic or Groq now raises ModelAPIError; Exa 402 errors are raised instead of retried

Breaking Changes

The release notes don't flag breaking changes, but there is one compatibility note:

• BrowserUse now runs on the host model by default, and browser-use telemetry is opt-in
  • Affected: projects relying on BrowserUse using a different model, or on browser-use sending telemetry by default
• WarnNearLimits removed from the Coder capability
  • Affected: harness setups that relied on near-limit warnings

Migration Guide

Upgrading from v2.55.0 to v2.56.0

pip install --upgrade "pydantic-ai==2.56.0"

After upgrading, check two things: wherever you use BrowserUse, confirm that running on the host model gives the cost and behavior you expect, and turn telemetry on yourself if you need it; wherever you pass MCPToolset(http_client=...) with your own auth, make sure auth isn't applied twice and that you now close the client yourself.

Compared With Other Frameworks

This release's microVM sandbox and the sandbox provider contract in same-day Mastra 1.77.0 tackle the same problem: the environment where agents run code has to be swappable and isolated. The difference is that Pydantic AI starts with a local microVM while Mastra starts with a provider abstraction over Docker and E2B.

Today's Takeaway

I assumed durable execution meant pairing with a standalone engine like Temporal. Pydantic AI adding a Postgres step store and then PGTaskDurability made me realize that for most teams, running durability on the Postgres they already have may be more practical than adopting a new engine.

References

• Pydantic AI v2.56.0 release notes — GitHub
• pydantic/pydantic-ai — GitHub
• SmolSandbox PR #10148 — GitHub
• PGTaskDurability PR #10174 — GitHub