Migrating Google Adk To Pydantic AI

Migrate Google ADK to Pydantic AI

Preserve caller-visible behavior, not ADK's class tree. Migrate the smallest complete request path and leave product infrastructure in the application.

Work from the running application

1. Read repository instructions, dependencies, tests, and runtime entrypoints. Record the installed Google ADK, Pydantic AI, and optional pydantic-ai-harness versions.
2. Trace one real request from Runner.run_async() or the deployed endpoint through the root agent, instructions, collaboration mode, model calls, tools, workflow nodes, callbacks/plugins, session service, emitted events, state/artifact deltas, and final response. Record only behavior that path uses.
3. Separate the contracts before choosing targets:
   • Session.events used as model context;
   • session-, user-, app-, and invocation-scoped state;
   • searchable long-term memory;
   • workflow/node checkpoints and resume IDs;
   • versioned artifacts and external side effects.
4. Classify the active slice:
   • Ordinary LlmAgent: normally one reusable Agent with typed dependencies, tools, and output.
   • ADK graph or dynamic workflow: keep simple deterministic control flow in plain async Python; use pydantic_graph when explicit typed nodes, branching, or graph inspection remain valuable.
   • Multi-agent delegation: inspect ADK's collaboration mode in Python 2.x; 1.x sub_agents use chat behavior. Use core multi-agent patterns for chat transfer semantics. Harness SubAgents can fit task or single_turn only when their isolated context, return, interaction, and concurrency behavior match.
   • Product runtime: retain auth, session/state stores, artifact stores, queues, transport, A2A endpoints, evaluation, observability, and deployment unless they are explicitly in scope.
5. Add deterministic characterization tests, then migrate one vertical slice behind its existing caller boundary.
6. Run the original tests and focused parity tests. Mark unexercised behavior unverified; similar names are not equivalence evidence.

Read Concept Mapping for the features the slice uses. Read Semantic Gaps for workflows, state, resume, confirmation, callbacks/plugins, event streams, skills, or execution environments. Read Verification and Cutover before removing ADK or changing production traffic.

High-risk gates

• Do not pass ADK Session objects or mutable state dictionaries through model-chosen tool arguments. Put authenticated identity and service clients in typed dependencies; persist product and workflow state through application-owned stores.
• message_history continues model context. It does not replace ADK session state, memory, artifacts, event records, node checkpoints, or invocation resume.
• ADK resumability replays recorded node/tool results and can execute tools more than once. Choose a durable execution design explicitly and prove restart plus idempotency for side effects.
• Keep conversational input, approval, and authorization distinct. Map tool confirmation to deferred tool approval; keep identity and access checks in trusted application code.
• Preserve callback/plugin ordering and short-circuit rules deliberately. Pydantic AI hooks have their own capability ordering and exception-based skip/recovery semantics; a list of lookalike hooks is not parity.
• ADK partial events are delivered without applying state deltas; each non-partial event applies its delta when appended. Artifact writes happen during the artifact operation, before the current event records the returned version. If callers consume ADK event fields or final-event detection, retain a boundary adapter and test the exact stream and persistence order.
• Use Harness only for an observed reusable capability. Ordinary agents need core only, and a command allowlist is not an OS security boundary.