Juan Carlos Orrego

[Juan Carlos Orrego @jcorrego]

7/ Don't automate the whole task. Share the actions.

Full argument → https://t.co/ZnDrjePKcs

[Juan Carlos Orrego @jcorrego]

1/ An agent is filling out a form. It pauses over the one obvious button, takes another screenshot, and I want to grab the mouse.

Two fields later, I want it to write the answer for me.

The unit of delegation is the action.

[Juan Carlos Orrego @jcorrego]

6/ I keep the default practical. Use machine interfaces for repeatable browser work and visual control where the DOM or API disappears.

But neither is enough until handoffs are explicit, authority is scoped, and resumption is designed.

[Ricardo Ruiz 🏛️ @RicardoRuiz_]

Presidente @petrogustavo, con mucho respeto se lo digo, porque me entristece que el jefe de Estado esté tan mal asesorado. Voy a explicarlo bien para usted y, sobre todo, para la ciudadanía. Irónicamente, esto sucedió cuando usted ganó, y de paso reviso el mismo puesto de votación en Suba y Villas de Granada, que cita el analista del cual erróneamente confía sumercé.

Lo que están mostrando es esto: en un puesto de votación con 24 mesas, las mesas 1 a 12 las gana Abelardo y las 13 a 24 las gana Cepeda, algo partido por la mitad y, según este señor, aparece en 1.350 puestos.

Visto así asusta. Y la observación es cierta: eso pasa, y pasa mucho.

Pero tiene una explicación que se puede comprobar, y voy a comprobarla con datos de la propia Registraduría. 🧵

[Juan Carlos Orrego @jcorrego]

7/ Automate every property you can state.
Keep humans where the question is whether you stated the right property — or whether the system still deserves to act.

Full argument → https://t.co/hg4nArrVm7

[Juan Carlos Orrego @jcorrego]

1/ The agent can generate more code than you can read.

So what happens to code review?

@unclebobmartin : stop reading every generated line; constrain the agent.
Dex Horthy: read the dang code.

They are both right — about different things. 🧵

[Juan Carlos Orrego @jcorrego]

6/ Still, humans own what automation cannot settle: intent, architecture, novel risk and irreversible consequences.

The code may be disposable. The codebase — its contracts, scars and reasons — is not.

[Addy Osmani @addyosmani]

Software quality now depends on the constraints you set around your agents.

When humans manually wrote most of the code we could look at the code itself for signs of quality. Is it clean? Is it thoughtful? Is it fast? Can another engineer understand it? Does it have tests?

Agents can now generate more code than people can read. When code generation scales beyond review, quality - checks for one or more of correctness, maintainability, security, performance etc - increasingly has to live somewhere else.

It moves into the harness, environment and operating system around the agent.

This can be the tests and deterministic checks that decide what the system is allowed to do (amongst others). Your constraints are what may eventually enable loops of agents to deliver production software reliably. They can include unit tests, property tests, acceptance tests, mutation testing and quality metrics.

This back-pressure lets the system resist bad work before it becomes somebody elses problem.

Set your constraints. They decide whether the code your agents generate is good enough to ship.

[Juan Carlos Orrego @jcorrego]

The likely future is convergence, not a victory parade.Full argument + the 5 questions I'd ask any team → https://t.co/82bfWivE38

[Juan Carlos Orrego @jcorrego]

Are LLMs a road to intelligence — or an off-ramp?

Team LLM: scale + next-token prediction keeps clearing every wall.

Team World Models: language is a shadow of reality, not reality.

Both are right. The fight is framed wrong. 🧵

[Juan Carlos Orrego @jcorrego]

Where I land: ship with LLMs for language, knowledge, code and agents.

Treat world models as the sharpest diagnosis of where they break: grounding, long-horizon planning, sample efficiency and learning from experience.