ricardo

sketchs of experiments.
testing agents, models, workflows.
notes when they’re ready. My views.
Lisbon

[ricardo @r_casaca]

https://t.co/4yGnghZ9hf

[ricardo @r_casaca]

What does a room learn the day power chooses silence over admitting a mistake?

Not the target. Everyone else. https://t.co/wCR6v0gMHR

[ricardo @r_casaca]

https://t.co/roM90hCKXu

[ricardo @r_casaca]

https://t.co/bPD7iRVZhh

[ricardo @r_casaca]

https://t.co/4h1P6pto8A

[ricardo @r_casaca]

Outsiders update their picture of you every time they watch you work. You never watch yourself work, you only watch the doubt on the way in, so your picture is still the one you set on arrival, against a room full of people who were already large. Being wrong about yourself upward is the one error everyone can see; being wrong downward is invisible and gets called humility, so the error you make is the one that costs you quietly. Other people sometimes profit from that, but the gap was yours before it was useful to anyone. Practice: ask three people who watched you work and have nothing to gain from your size what they saw, and take their answer over your own.

[ricardo @r_casaca]

https://t.co/Pz9VN5GzPW

[ricardo @r_casaca]

https://t.co/Pz9VN5GzPW

[ricardo @r_casaca]

Sounds like your friend isn't stuck because the model is weak, he's stuck because the job is still write the plan, cut the epics, wait for the nod, and that "furniture" was there long before the model arrived. The difference now is that an agent won't nod it through. You have to write each step out, and the ones you can't write down were never steps, probably they were people...

[ricardo @r_casaca]

@jeffteper Cost per value is the quiet one... Still plenty of road ahead.

[ricardo @r_casaca]

TLDR — The communication problem
We have a communication problem" is a sentence with no subject — until power sits down. https://t.co/iJGU0Ip0aU

[ricardo @r_casaca]

OpenAI shipped an Agents API a few days ago. Keeping an agent running for days, remembering where it was, recovering when a step fails: all of that is now something you rent. The launch post has over three thousand bookmarks, and underneath it the people who run two harnesses side by side are arguing about something else entirely, which one pauses to ask before it acts, and which one is powerful in a chainsaw kind of way.

I run agents from three vendors on my week, every day. The rules they run under fit on one page, and the page says the same thing for all three: write, log, never publish, never delete, never decide.

That's not caution. It's the split. Hours alone is now the benchmark, and rented by the token. Decisions allowed is the job, and nobody rents it.

The week nothing moved on my desk, it wasn't the agents' permissions that were missing. It was my decisions.

[ricardo @r_casaca]

https://t.co/IjyJsyj7XP

[ricardo @r_casaca]

https://t.co/IjyJsyj7XP

[ricardo @r_casaca]

@alex_holovach perfect timing for me to 🫡

[ricardo @r_casaca]

https://t.co/fKmzR3YMqe

[ricardo @r_casaca]

Agents gave a perfect backlog. It didn’t move.

Visibility isn’t scarce anymore. What’s left is drop + whose name. A backlog is not a decision.

[ricardo @r_casaca]

https://t.co/YwZqZ3GgGq

[ricardo @r_casaca]

https://t.co/YwZqZ3GgGq

[ricardo @r_casaca]

@Jason_Stapleton @satyanadella Yes, every 1st day of the month, then get out of credits and switch to grok, claude, codex...

[ricardo @r_casaca]

In case you missed it Grok Bot is now live on Android. Finally

https://t.co/FshVizqkcI

[ricardo @r_casaca]

https://t.co/05TPgSiwCM

[ricardo @r_casaca]

The useful split isn't "AI that codes" vs "AI that chats". It's one bot that owns judgment, and a factory floor you can actually watch. I don't want eight agents inventing work. I want a chief of staff and a pane I can see.