migrating-langchain-to-pydantic-ai

Trace one representative request through prompts, tools, state, and streaming, then rebuild the smallest behaviorally complete slice as a pydantic_ai.Agent with typed dependencies and tools. Preserve public request and response shapes with a small adapter and prove parity with characterization tests.

Pydantic AI message_history only provides conversation continuity, not checkpoint replay, forks, or pending writes. Split state into dependencies, messages, workflow state, and long-term memory, then use application persistence or a durable execution integration such as DBOS, Temporal, or Prefect.

Yes, tool_from_langchain and LangChainToolset wrap existing LangChain tools as a transitional bridge. The wrapper keeps LangChain validation and dependencies and does not preserve flags like return_direct, so add parity tests and plan removal.

No. run_stream() commits the first matching streamed output immediately, so co-emitted tools and retries can produce a different terminal result than run(). Use run(event_stream_handler=...), run_stream_events(), or iter() when complete tool execution must finish before the final result.

Do not use it for migrations centered on create_deep_agent or Deep Agents harness features such as planning, skills, filesystem, subagents, or sandbox contracts. Those are harness migrations that require different scope decisions.

Classify every observed contract as verified-equivalent, verified-adapter, intentional-change, external-owner, not-applicable, unverified, or blocked, each linked to an executable test. Use Logfire traces as supporting evidence only, and remove LangChain dependencies only after no retained path needs them.