Pydantic AI v2.63.0

Summary

Pydantic AI 2.63.0 was released on 2026-10-27, one day after 2.62. It adds a capability that stops agents stuck in repeated-action loops, an IPython-based code mode, shared spend budgets in PostgreSQL, and per-task timeouts for evals. It also supports GPT-6.1 Terra and AG-UI protocol 1.0.

Key points

• DetectLoops: a capability that interrupts an agent when it keeps repeating the same action.
• Code mode: new IPythonCodeMode next to the Monty-based CodeMode. CodeMode now defaults to dynamic_catalog=True, so newly revealed tools are appended instead of rewriting the cached prompt (better prompt-cache hits).
• Budgets: PostgresSpendStore lets several processes share SpendLimits budgets.
• Evals: Dataset.evaluate accepts task_timeout to bound each task attempt.
• LLMReminder now runs every 10 requests by default instead of on every request.
• UI adapters: manage_message_history stops server-held history from being sent to the model twice. AG-UI protocol 1.0 is supported.
• Structured output: strict schemas are kept within Anthropic and OpenAI complexity limits, falling back to non-strict output when needed.
• New public pydantic_ai.framing module (wrap, neutralize_tags, register_tag) for wrapping untrusted text in tags.
• Upgrade with pip install -U "pydantic-ai==2.63.0".

Why it matters

Loop detection and shared spend limits address two common failure modes of unattended agents: getting stuck and running up cost.