Juan Gonzalez

Co-founder & Lead AI Engineer @sancrisoft
Shipping AI-native software & agentic systems in production | AI-native product engineering

[Aníbal Rojas @anibal]

Esta ¿advertencia? de Satia Nadella sobre el impacto de la Inteligencia Artificial nace, se apalanca y es central al desarrollo de software. Y la realidad es que si los grandes laboratorios no solo crean modelos cada vez mejores, sino que capturan el "institutional knowledge" de grandes verticales de la industria (vean nuestro episodio sobre los FDEs)... Pues todo se puede ir a la carajo porque al final hace falta una economía, pero... esa es mi opinión, vamos a ver qué piensa @fmontes y sobre todo qué piensan ustedes sobre esta advertencia/blueprint del CEO de Microsoft. Como siempre aquí en vivo todos los lunes en la tarde/noche de Latinoamérica. Vayan a ocorres punto com y ahí tienen todos los enlaces para los diferentes streams y pueden agregarlo al calendario para participar en vivo con sus puntos de vista en esta discusión sobre nuestro futuro.

[Juan Gonzalez @_juangonz]

The paradox nobody expects: adding more AI agents to your dev workflow gives humans MORE control, not less.

Each agent stops at a checkpoint with a reviewable artifact. Compare that to one dev self-reviewing their own assumptions at every step. Which has more oversight?

[Amazon Web Services @awscloud]

More AI-generated code doesn't make your team faster. It might actually slow you down.

[Juan Gonzalez @_juangonz]

Five agents, five artifacts, five human checkpoints. The AI produces the work; engineers own the judgment. Full breakdown: https://t.co/t7q253xmRZ

[Juan Gonzalez @_juangonz]

We stopped treating AI as autocomplete and built five specialized agents that mirror a real engineering team. Each owns one role. Each stops for human review before the next phase begins. Here's how the five agents work:

[Juan Gonzalez @_juangonz]

5/ The QA agent. No freestyle testing. It works from a risk register where every test traces to a risk and every risk traces to a user story. You can finally answer "are we testing the right things?" with data.

[Matt Pocock @mattpocockuk]

I poured my 10 years of teaching experience into a skill.

It's called /teach, and it can teach you anything.

Here's how it taught me to solve a Rubik's cube:

[Juan Gonzalez @_juangonz]

Give one AI agent all the responsibility — spec, design, code, review, tests — and it optimizes for the path of least resistance: straight to code.

Same as humans without role clarity. When everyone owns everything, no one is accountable for anything.

[Juan Gonzalez @_juangonz]

This morning you wrote the spec. By lunch you designed the schema. Afternoon: code. 4 PM: reviewed your own PR. 6 PM: shipped and hoped.

Being PM, architect, dev, reviewer, and QA at once isn't a badge of honor. It's where tech debt is born.

Which role do you skip first?

[Juan Gonzalez @_juangonz]

The deepest shift in agentic engineering isn't speed.

When AI handles "did we build it right?", covering clean code, style, and tests, humans get bandwidth for the harder one: "did we build the right thing?"

That's where the value lives. AI can't answer it.

[Juan Gonzalez @_juangonz]

@anibal Que trauma tenemos 😢 fantástico ver los boletines online y el 99% listo en menos de 2h.

[Juan Gonzalez @_juangonz]

Architecture decisions lived in two places: someone's head, and a Slack thread no one would find again.

When the engineer leaves, the rationale evaporates.

ADRs were always the answer. Agentic systems finally make them tractable on every feature.

[Every 📧 @every]

@claudeai Vibe Check! Opus 4.8 tops both our Senior Engineer benchmark and our AI writing tests. It’s the most complete model we’ve tested: https://t.co/qp5qifN8tw

[Juan Gonzalez @_juangonz]

AI handles structured execution. Humans hold the decisions.

That's the whole shift. Everything else is implementation detail.

https://t.co/eT4ldWSrFW

[Juan Gonzalez @_juangonz]

Every agentic system needs explicit checkpoints, moments where AI work pauses and waits for a human.

Not bureaucratic gates. Decision points.

We use five. Here's the full framework.

[Juan Gonzalez @_juangonz]

The reason it works isn't the count. It's the context the reviewer has at each step: spec, architecture doc, risk register, all connected.

The engineer isn't pattern-matching syntax. They're evaluating whether the whole chain holds together.