Pular para o conteúdo
TakoAPI
Voltar às skills

agent-evaluation

por rustyorb

DevOps & Cloud
2.6kDownloads informados pelo clawskills.sh
5Estrelas
0Visualizações
0 curtidaEntre para curtir esta skill
—Ainda sem avaliações
Sua avaliação

Resumo

Testing and benchmarking LLM agents including behavioral testing, capability assessment, reliability metrics.

Descrição

O que esta skill faz

Evaluating LLM agents requires different approaches than traditional software testing because the same input can produce different outputs. It covers behavioral regression tests, capability assessments, and reliability metrics designed to catch issues before production. Even top agents score below 50% on real-world benchmarks.

Standard unit testing misses the probabilistic nature of LLM outputs, so statistical and behavioral approaches are needed to get meaningful reliability signals.

Quando usá-la

  • Catching agent regressions before deploying a new model version
  • Measuring how reliably an agent handles ambiguous or edge-case inputs
  • Designing benchmarks that reflect actual production conditions
  • Detecting when an agent optimizes for a metric instead of the real task
  • Preventing test data from leaking into agent training or prompts

Exemplo de fluxo de trabalho

ENTRADA

User asks: evaluate the reliability of my customer support agent on ambiguous queries

AGENTE
  1. 1

    Defines behavioral invariants the agent must always satisfy, such as never revealing PII

  2. 2

    Runs each test case multiple times to build result distributions and flag flaky behavior

  3. 3

    Generates adversarial inputs designed to break or confuse the agent

  4. 4

    Calculates reliability metrics across runs and compares them against a baseline

  5. 5

    Reports which invariants passed, which failed, and which showed statistical instability

SAÍDA

Evaluation report with pass rates, flakiness scores, and specific failure cases flagged for remediation

Instalação

clawhub install rustyorb/agent-evaluation

Execute o comando da CLI ou cole o prompt no Claude Code, Cursor, Codex ou no agente de código que você usa.

Links