convez.ai
AgenteDevOps & InfraestruturaAvançado
tomatefarcie123

Agentboarding

tomatefarcie123/agentboarding

CI para a era dos agentes: testa adoção de produtos por IA, autopsia falhas, auto-corrige

Ver no GitHub
InicialPluginScore · de 0 a 100

CI/CD para agentes de IA que testa automaticamente se um agente Claude consegue adotar seu produto usando apenas a documentação pública, identifica falhas específicas (docs ambígua, flags alucinados, resumos incompletos), auto-corrige e reprova em sandbox. Transforma fracassos invisíveis de agentes em métricas visíveis e corrigíveis.

Instalação

$npx convez-ai tomatefarcie123/agentboarding

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Testar se agentes Claude conseguem integrar seu produto com docs existentes
  • Identificar gaps em documentação que frustram agentes de IA
  • Automatizar detecção e correção de problemas de UX para agentes
  • Medir e melhorar AX Score (Agent Experience) de ferramentas

Veredito

Ferramenta experimental sofisticada (Claude Build Day 2026) que redefine QA para a era de agentes de IA. Aborda um problema real: agentes abandonam produtos silenciosamente quando ficam presos. Implementação em sandbox com auto-fix é elegante, mas requer infra específica (gmsaas, adb, Claude Code) e mentalidade de DevOps avançada.

README

A tradução para o português deste README ainda está na fila.

Agentboarding

CI for the agent era. Built at Claude Build Day (Shack15 SF, June 2026).

AI coding agents like Claude Code are becoming the buyers and integrators of developer tools: a developer asks an agent "how do I use X?", and the agent reads the docs and tries to set it up itself, in the terminal — no website visit. When it gets stuck, it quietly gives up and picks a competitor, and that failure is invisible today.

Agentboarding makes it visible, reproducible, and fixable. It turns a real, sandboxed Claude Code agent loose on a goal — "from a clean machine, boot a cloud Android device with Genymotion" — using only the product's public docs, records every step, root-causes each failure (product broken / docs ambiguous / hallucinated flag / summarization casualty / funding wall), auto-fixes the gap, and re-runs a fresh agent to prove it closed — an AX Score moves from red to green.

Layout

PathWhat
agentboarding/orchestrator, sandbox + bash gate, trajectory recorder, HaikuShred replay, root-cause classifier, auto-fix generator, AX Score, x402 funding rail
tests/ + tests/golden/pytest suite + recorded golden trajectories
geo/, analytics/support files reused from the Rozz GEO stack (Schema.org hardener, AI-bot traffic classifier, the gmsaas runbook fixtures)
hackathon/build-day/the brief, rubric, test-suite spec, and demo runbook/script that directed Claude

Run the tests

python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python -m pytest -q

Live agent runs additionally need these on PATH: the claude CLI, adb (Android platform-tools), and node (Turndown). The agent installs gmsaas itself — it's the product under test, not a prerequisite. (gmsaas is only needed on the host for the operator's funding pre-flight, funding.preflight_funding.)

What "clean machine" means (scope). A representative target environment: the base Android toolchain (adb, node, python) present, the product absent. The agent owns everything product-specific — install gmsaas → auth (without leaking the token) → config set android-sdk-path → boot → adbconnect → adb shell. Generic tooling the runbook assumes (adb) is part of the environment on purpose: Agentboarding scores what the vendor can fix from their docs, not the agent's ability to install an OS-level SDK.

How it was built

Authored by Claude Opus 4.8 in Claude Code: a multi-agent workflow produced the build spec — goals, evals, a 100-point rubric, and this test suite (see hackathon/build-day/BUILD_PROMPT.md) — which a headless Claude Code session then implemented block-by-block against the suite. Demo walkthrough: hackathon/build-day/DEMO_SCRIPT.md.

Ver o original no GitHub

Bom para

Relacionados