An OpenAI agent finds the weather for a US zip code. K9-AIF governs every step it takes: who may ask, what reaches the model, each tool call and result, every reply, and which model does the thinking.
demo / demo · your session is private to youA well-made library for agent loops. From its documentation:
One configuration that holds for every agent on every runtime:
The adapter is a K9-AIF orchestrator wrapped around the SDK. Each numbered gate is a K9-AIF checkpoint on a real crossing.
The SDK does the work. K9-AIF decides at each crossing.
Recorded with ./run.sh --demo against a local model. In the demo, every request shows
a receipt like this.
The adapter is one more orchestrator. It extends the same
BaseOrchestrator contract as every K9-AIF orchestrator, so the Router sends it work like any other.
On this path the SDK's Runner and agent take the place of Squad and Agent, while Shield, Zero Trust and the
Model Router stay shared with every other path.
The K9X ecosystem brings the framework to life: K9X SATAN fires attacks at Shield to prove it holds, K9X HIL handles human decisions, K9X Studio builds K9-AIF applications visually, and DAS is a full reference application.