본문으로 건너뛰기
TakoAPI
스킬로 돌아가기
—아직 평가 없음
내 평가

요약

Ship agents that fix themselves.

설명

이 스킬이 하는 일

Kalibr routes AI agent execution paths based on real production telemetry. You define candidate paths (model + tools + parameters) and it learns which performs best for each task, shifting traffic away from failures or degraded providers automatically. Uses Thompson Sampling with 10% canary traffic to catch degradation before users notice.

Unlike LiteLLM or LangChain fallbacks that react to failures after the fact, Kalibr routes proactively based on learned success rates per task, so a degraded provider is avoided before a request is lost.

사용 시점

  • Automatic failover when OpenAI or Anthropic has an outage
  • Routing different agent tasks to different models based on past success rates
  • Reducing LLM costs without manually tuning model selection
  • Detecting when a model starts returning worse results over time
  • Replacing a hardcoded model name with adaptive path selection

예시 워크플로

입력

User asks: Route my email extraction agent across GPT-4o, Claude, and Gemini and adapt based on what works

에이전트
  1. 1

    Import Router from kalibr and define three candidate paths with different models and parameters

  2. 2

    Call router.completion() with the user message, which Kalibr routes to the current best path

  3. 3

    Evaluate the response and call router.report(success=True/False) or use success_when= for auto-reporting

  4. 4

    Kalibr updates internal telemetry and adjusts path weights after each outcome

  5. 5

    After ~50 outcomes, routing is locked in and continuously adapting via canary traffic

출력

Agent calls automatically route to the highest-performing model path, with degraded paths deprioritized without manual intervention

요구 사항

  • KALIBR_API_KEY environment variable
  • KALIBR_TENANT_ID environment variable
  • Kalibr account at https://dashboard.kalibr.systems/settings

설치

clawhub install devonakelley/kalibr

CLI 명령을 실행하거나, 프롬프트를 Claude Code, Cursor, Codex 등 사용하는 코딩 에이전트에 붙여넣으세요.

링크