Agents.KT
A typed Kotlin DSL runtime for building AI agent systems on the JVM, where the compiler enforces input/output types and tool permissions, and every run produces a tamper-evident audit trail.
Agents.KT is a typed agent runtime for the JVM built for teams that need to state exactly what an AI system is allowed to do. Every agent is typed with one input type and one output type; type mismatches and invalid compositions are caught by the compiler because composition is purely type-driven, and structural misuse fails fast at construction time rather than at run time.
An agent's skills are either a deterministic Kotlin lambda with no LLM involved, or an agentic skill where the LLM picks from an allow-listed set of tools that the runtime refuses to let it call outside of. The 0.6-0.7 release line adds deterministic permission manifests, a runtime manifest-hash correlation, JSONL audit export, a tamper-evident audit ledger, OpenTelemetry/LangSmith/Langfuse bridges, before-interceptor policy hooks, vendor-neutral prompt caching, and snapshot/resume with a manifest-hash restore guard.
It requires JDK 21+ and Kotlin 2.x, and states plainly that it is not a compliance product and does not OS-sandbox arbitrary tool code.