OpenAI Agents SDK vs Google ADK

Few primitives versus a workflow engine

OpenAI states its design principle as having enough features to be worth using but few enough primitives to be quick to learn. An agent is a model with instructions and tools; a handoff transfers control to another agent; a guardrail validates input or output; a session keeps history. Orchestration is ordinary Python or TypeScript. Realtime and voice agents are part of the same package.

ADK was always more structured, and version 2.0 made the structure explicit. Google's documentation says 2.0 moves from hierarchical execution to a graph-based architecture, with graph workflows for deterministic routing, dynamic workflows for loops and branching in code, and collaborative workflows for coordinator agents with subagents. The Python SDK reached general availability on 19 May 2026, Go on 30 June and TypeScript on 21 August. The migration notes list breaking changes: new fields in the event schema, agents executing as nodes so that legacy execution hooks are bypassed, and a warning that appending events to a session by hand is unsafe.

Sandboxes and state

OpenAI's sandbox agents are the more distinctive feature of the pair. The documentation describes a persistent workspace in which the model can search document sets, edit files, run commands and resume from saved sandbox state. A manifest stages files and mounted directories; sessions and snapshots let a later run reconnect to earlier work; capabilities include shell, filesystem editing, skills and compaction. Developers choose where the workspace lives. The Unix-local client is for trusted development, and the docs note it provides no network isolation on Linux; Docker provides container isolation; seven hosted providers have named clients. This is the same separation of harness and execution environment described on the agent sandbox page.

ADK's documentation emphasises session, state and memory management and what it calls treating context like source code. Recent releases added a SQLite memory service, tool confirmation inside workflows, graceful cancellation and automatic model failover. A sandbox abstraction comparable to OpenAI's is not documented in the ADK pages reviewed for this comparison; on Google Cloud, code execution is offered by the platform, not the SDK.

Model neutrality in practice

Both vendors claim openness and both qualify it. OpenAI's repository calls the SDK provider-agnostic, but its models guide lists what degrades with other providers: tracing needs an OpenAI API key unless disabled or redirected, many providers lack the Responses API so calls fall back to Chat Completions, structured outputs may not be schema-validated, and hosted tools such as web search and file search are unavailable. Tracing is also unavailable to organisations on a Zero Data Retention policy.

ADK's repository describes it as model-agnostic and deployment-agnostic while "optimized for Gemini". Release notes show active work on other vendors' models, including visible thinking for Claude models in v2.9.1 and OpenAI reasoning-model support in v2.10.0. Neither vendor publishes a compatibility matrix that would let a buyer verify parity across models in advance, so cross-model behaviour should be tested, not assumed.

Evaluation and deployment

ADK includes evaluation in the framework. It compares an agent's tool-call trajectory with an expected one and scores the final response, using test files for single interactions and evalsets for multi-turn sessions, with a CLI, a web UI, pytest integration and simulated users. The OpenAI SDK's equivalent built-in facility is tracing, which records generations, tool calls, handoffs and guardrails and exports them to OpenAI's dashboard or to third-party processors; scoring is left to other tools. See agent evals for why trajectory-level grading matters.

For deployment, ADK documents four routes: Google's fully managed Agent Runtime, Cloud Run, GKE and any container infrastructure. The OpenAI SDK is a library that runs inside the developer's own service. OpenAI's hosted alternative is the separate Agents API, in beta, where OpenAI runs the harness; that trade-off is covered in managed agents vs self-hosted agent harnesses.