Koog 0.4.0, an replace to the JetBrains Kotlin framework for constructing AI brokers, options native structured output meant to carry up in manufacturing. It additionally provides Apple’s iOS as a goal platform, in addition to GPT-5 mannequin and OpenTelemetry help.
Introduced August 28, Koog 0.4.0 is meant to make brokers observable, seamless, deployable in a consumer’s stack, and extra predictable, whereas including help for brand new platforms and fashions, JetBrains mentioned. Code is accessible from GitHub. Native structured output addresses a scenario wherein a giant language mannequin (LLM) can present the precise information format wanted, however then the method stops working. Koog 0.4.0 provides native structured output supported by some LLMs, with pragmatic guardrails like retries and fixing methods. When a mannequin helps structured output, the framework makes use of it instantly. In any other case, Koog falls again to a tuned prompt-and-retry, with a fixing parser powered by a separate mannequin, till the payload appears to be like precisely the way it ought to.
For iOS, Koog is now obtainable on Apple’s cell working system as a part of a give attention to Kotlin Multiplatform. Builders can construct an agent as soon as and ship it to iOS, Android, and JVM again ends—all with the identical technique graphs, observability hooks, and assessments. However builders should use Koog 0.4.1 to construct for iOS. Moreover with Koog 0.4.0, GPT-5 help and customized LLM parameters assist a mannequin suppose more durable on advanced parameters. Settings akin to reasoningEffort allow a steadiness of high quality, price, and latency for every name, based on JetBrains. OpenTelemetry backing, in the meantime, helps each the W&B Weave AI improvement toolkit and the Langfuse open supply LLM engineering platform.