Show HN: Kybernis – AI 에이전트가 동일한 작업을 두 번 실행하는 것을 방지합니다.
hackernews
|
|
🔬 연구
#ai 에이전트
#anthropic
#api 비용 절감
#review
#sdk
#오픈소스 cli
#중복 실행 방지
원문 출처: hackernews · Genesis Park에서 요약 및 분석
요약
AI 에이전트가 실제 시스템을 제어하며 발생하는 비동기 작업의 재시도 과정에서 중복 실행으로 인한 결제 오류나 데이터 불일치 문제가 발생하고 있습니다. 이를 해결하기 위해 개발된 'Kybernis'는 에이전트와 외부 API 사이에서 실행 의도를 기록하고 멱등성을 보장하는 신뢰성 계층을 제공합니다. LangGraph나 AutoGen 같은 주요 프레임워크와 호환되며, 생산 환경에서 변형 작업이 정확히 한 번만 실행되도록 보장하여 안정성을 확보합니다.
본문
Stop your AI agents from double-charging APIs. Diagnose agent failure loops locally with our open-source CLI, or drop in the SDK to hard-block semantic double-spends before they hit production. LLMs don't understand execution boundaries. Network lag causes timeouts. Timeouts cause LLMs to retry. If your agent times out on a payment API, it just issues a second payment. This is the Semantic Double-Spend. Without Kybernis Agents retry blindly on errors, causing duplicate mutations in your database. With Kybernis We evaluate semantic intent and hard-block hallucinated retries instantly. Observability X-Ray your local agents. Watch them attempt duplicate writes before you deploy. Get started in 2 minutes No complex dashboards. Just code. Install SDK Install the package. npm install @kybernis/sdk Wrap your Tools Wrap any LangGraph or MCP tool. import { withGuard } from '@kybernis/sdk'; const safeRefund = withGuard(paystackRefund); Execute Safely Kybernis guarantees exactly-once execution. await safeRefund({ amount: 50000 }); Developer FAQ Because engineers hate surprises. No. The Kybernis SDK evaluates the semantic signature locally in under 10ms before allowing the network request to proceed. It operates entirely out of the critical path of your external API response. Absolutely not. We hash the execution context and semantic intent into a secure signature locally on your machine. We never see or store your raw PII, API keys, or payload data. No. Kybernis is an execution safety layer, not an orchestrator. You still build your agents in LangGraph or connect them via Anthropic MCP. You simply wrap your high-risk tools with the Kybernis SDK to protect the execution boundary. Built for developers who hate incident reports. We kept running into the same problem: putting agents in production was terrifying because of unconstrained retries. So we started building Kybernis—a faster, safer way to orchestrate tool execution. Kybernis is actively used in our internal agent workflows. We are onboarding 15 developers this month to shape the product. Join 17 other devs currently testing the CLI. Takes 2 minutes. No setup. No BS.
Genesis Park 편집팀이 AI를 활용하여 작성한 분석입니다. 원문은 출처 링크를 통해 확인할 수 있습니다.
공유