ia-engineer-framework

Framework open-source para hacer agentes IA fiables en produccion. 6 tiers progresivos: desde controlar tu agente en 10 minutos hasta CI/CD con evaluaciones automaticas.

Open-source framework for making AI agents reliable in production. 6 progressive tiers: from controlling your agent in 10 minutes to CI/CD with automated evaluations.

GitHub Quick Start

El problemaThe problem

Especificaciones vagasVague specifications
El agente solo puede adivinar lo que quieresThe agent can only guess what you want

Sin verificacionNo verification
Dice "listo" sin ejecutar testsSays "done" without running tests

Sin estadoNo state
Cada sesion empieza de ceroEvery session starts from scratch

Sin observabilidadNo observability
No sabes cuanto cuesta ni si la calidad bajaYou don't know the cost or if quality drops

Sin CI/CDNo CI/CD
Cambios del agente llegan a produccion sin verificarAgent changes reach production unverified

Convenciones no escritasUnwritten conventions
El agente no puede cumplir lo que no conoceThe agent can't follow what it doesn't know

6 tiers progresivos6 progressive tiers

Harness

CLAUDE.md, AGENTS.md, init.sh, feature_list.json, progress.md, session-handoff.md

10 min

Eval

4 datasets JSONL (32 escenarios), run-evals.py, score-evals.py, ci-gate.py, rubrica 6 dimensiones

4 JSONL datasets (32 scenarios), run-evals.py, score-evals.py, ci-gate.py, 6-dimension rubric

30 min

Observe

Logger 10 campos, cost tracker, dashboard Grafana, guia PostHog, alertas presupuesto y drift

10-field logger, cost tracker, Grafana dashboard, PostHog guide, budget and drift alerts

Hooks

Pre-commit lint, post-task verify, context check, cost guard ($10/session), security scan

15 min

Patterns

Diagnostic loop, multi-session, graph engineering, circuit breaker, HITL, model routing, prompt versioning, agent testing

lecturareading

CI/CD

GitHub Action para evals en PR, quality gate, pre-push check

GitHub Action for PR evals, quality gate, pre-push check

Empieza en 5 minutosGet started in 5 minutes

git clone https://github.com/pedri77/ia-engineer-framework.git

cp -r ia-engineer-framework/harness/ your-project/

nano your-project/CLAUDE.md
nano your-project/feature_list.json

# Run your AI agent

ComparativaComparison

	Learn Harness Eng (11.3K stars)	ia-engineer-framework
Scope	Solo harnessHarness only	Harness + eval + observe + CI + patterns
Eval pipeline	No	Runner + scorer + CI gate
ObservabilidadObservability	No	Logger + cost tracker + alerts
CI/CD	No	GitHub Action + quality gate
IdiomaLanguage	InglesEnglish	Espanol nativoSpanish native
FormatoFormat	CursoCourse	Copy-paste to production
Lock-in	EspecificoSpecific	Claude Code, Codex, Cursor

PrincipiosPrinciples

Automatizacion > IAAutomation > AISi un test lo resuelve, no uses LLMIf a test solves it, don't use an LLM

Verificable > inteligenteVerifiable > smart8 evals > 1 prompt perfecto8 evals > 1 perfect prompt

Copy-pasteCada archivo funciona standaloneEvery file works standalone

ProgresivoProgressiveTier 1 en 10 min, Tier 6 con equipoTier 1 in 10 min, Tier 6 with a team

MedibleMeasurableSi no mides, no implementesIf you can't measure it, don't ship it

Sin lock-inCualquier agente IAAny AI agent