Kotlin AI agents that survive crashes and LLM swaps

What it does

Koog is a Kotlin Multiplatform framework for constructing AI agents through a type-safe Kotlin DSL that also speaks Java. It orchestrates calls to OpenAI, Anthropic, Google, DeepSeek, Ollama, Bedrock, and other providers, giving agents tool use, streaming responses, graph-based workflows, and vector-memory retrieval. The framework embeds into Spring Boot or Ktor applications and emits traces to observability backends such as Langfuse and W&B Weave.

The interesting bit

Koog treats LLM provider lock-in as a runtime annoyance rather than an architectural constraint: you can reroute an ongoing conversation to a different model or provider without losing history, and it compresses that history intelligently to keep token usage in check. That combination of portability and thrift is unusual in the current agent-framework landscape.

Key highlights

• Targets JVM, JS, WasmJS, and iOS.
• Built-in retries and agent persistence to restore state after failures.
• Integrates MCP tools and supports the Agent Client Protocol (ACP).
• Provides vector-embedding-based RAG and memory across conversations.
• Ships with OpenTelemetry exporters for tracing and monitoring.

Caveats

• The README lists Android among deployment targets in the feature overview, but the supported-targets section currently omits it, so mobile support appears partially inconsistent.
• It carries a JetBrains incubator badge, so long-term commitment may differ from flagship products.
• The koog-agents-additions module is at 1.0.0-beta, suggesting some capabilities are still settling.

Verdict

Teams already invested in Kotlin Multiplatform or JVM backends who need structured, observable agent workflows should evaluate this. If you are not running in the Kotlin or Java ecosystem, the value proposition drops sharply.