Marcos Lourenço @mvlcode

Head of Engineering, 18 years exp. AI agents in a 40-repo production codebase + side projects. The cost, the hit rate, and the parts that break 🇧🇷🇳🇱🇪🇸🇺🇸 Spain Joined September 2025

Tweets

128
Followers

45
Following

53
Likes

81

Marcos Lourenço @mvlcode

3 weeks ago

@SandeepSan69801 @AnthropicAI @OpenAI I normally work with work items, and in my instructions I ask it to update it as it goes. So if I need to resume the session, later, or another day, I just provide the work item id. I use Azure DevOps, but any board tool like ClickUp, Notion could work providing a PAT.

0 0 1 29 0

View Details

Marcos Lourenço @mvlcode

3 weeks ago

Yea that's def true. That's why we "humans" are still needed in the loop. More and more the hard part is done by AI, but the last bits need to be done by us. A great thing to help reaching that finishing line is combining agents with other capabilities, such as testing, and having it giving it back to the implementation agent with evidences to fix it. However, there are always findings and blockers on the way only you can decide on. Unless you create a third agent that is going to make decisions on your behalf, but the outcome might not be the desired one 😅

0 0 0 9 0

View Details

Marcos Lourenço @mvlcode

3 weeks ago

@andrevalentim_ Parabens. Como voce tem monitorado a performance e a usabilidade do resultado? Voce deixa eu agente de forma independente endereçar isso?

0 0 0 36 0

View Details

Marcos Lourenço @mvlcode

3 weeks ago

Same page, 50x faster. 2.00s becomes 0.04s, and the counting happens once a night instead of once per visitor. Whatever you are building, ask your AI tool one question before it writes the code: does this have to be live? Most of the time the answer is no. Built with Claude Code by @claudeai.

0 0 1 21 0

View Details

Marcos Lourenço @mvlcode

3 weeks ago

Your app probably does the same expensive work over and over. A voting results page that counts 2 million votes on every visit takes two seconds. Every visitor pays for it, and they all get the same answer. Most data does not need to be real time. Count it once at 3 AM, save the answer, let the page read that. Same page, 50x faster. Ask your AI tool one question before it builds: does this have to be live?

5 0 3 101 0

View Details

Marcos Lourenço @mvlcode

3 weeks ago

Most data does not need to be real time. If your numbers can be a few hours old, do the counting once at 3 AM, save the answer, and let the page read that instead.Every 6 hours for a poll that already closed. Every 3 hours for a trend. Every hour for a board people watch move.

0 0 1 7 0

View Details

Marcos Lourenço @mvlcode

3 weeks ago

Why is it slow? It counts all two million votes from scratch, every time anyone opens the page. Nobody asked a different question. Everyone gets the same answer, and everyone pays for it again.

0 0 1 6 0

View Details

Marcos Lourenço @mvlcode

3 weeks ago

Your app probably does the same expensive work over and over. Here is a voting results page that counts 2 million votes on every single visit. Two seconds, every time. Then the same page after one small change. Watch the stopwatch 👇

0 0 1 9 0

View Details

Marcos Lourenço @mvlcode

3 weeks ago

@digoabreu verdade, marketing digital também é outra skill muito importante hoje em dia.

0 0 0 13 0

View Details

Marcos Lourenço @mvlcode

3 weeks ago

@naveen_saradhi_ what about interacting with AI to design the schema the you want, other than doing manually? ofc you can do it manually, but feels more like a hobby other than strictly necessary.

0 0 0 2 0

View Details

Marcos Lourenço @mvlcode

3 weeks ago

@tchaguitos Boa! Tambem tenho feito isso, o unico problema é quando tem muito PR pra validar né? Ou quando agente volta com alguma pergunta mid-flight 😅

0 0 0 9 0

View Details

Marcos Lourenço @mvlcode

4 weeks ago

@juanquijandria8 as long the baby sitting part is well paid 🤣

0 0 1 36 0

View Details

Marcos Lourenço @mvlcode

4 weeks ago

Def true, and I would add the boring half of why it works. These past weeks I had Claude map out the UI test scenarios on a project and run them. It comes back with the scenario that failed instead of me guessing from a stack trace. That is what actually changed for me. Not that the agent writes better code, but that I find out faster when it did not. The prod horror stories almost always have that step missing. I would love to know what Stripe has sitting underneath that 1,300 number.

1 0 1 50 0

View Details

Marcos Lourenço @mvlcode

4 weeks ago

Uso Claude Code todo dia e mantenho as specs como cards, porque pedido vago faz ele ler o repositório inteiro antes de começar a trabalhar. Preparo o card com outros modelos e deixo o Fable para implementação. Isso tem me ajudado a nao exceder o limite. Alem de ter algumas skills com contextos especificos de partes do sistema.

0 0 1 21 0

View Details

Marcos Lourenço @mvlcode

4 weeks ago

@gabjeesus Parabens! 🚀

0 0 0 13 0

View Details

Marcos Lourenço @mvlcode

4 weeks ago

@abaderenan 😂

0 0 0 3 0

View Details

Marcos Lourenço @mvlcode

4 weeks ago

Same here, I was using WisprFlow for a while. Its nice that it can integrate with everything. But I also just feel like typing still much better. I think a nice use case would be, if you are in the middle of a task, and the agent asks you in voice mode what you want to do about a mail, or a notification you just received. Then you can easily reply back, without completely shifting your attention to where it came from. Like "Team Member A sent you a message about X", "Ok, tell you him he can find it at Y" I don't believe too much in using for every day's tasks.

0 0 0 48 0

View Details

Marcos Lourenço @mvlcode

4 weeks ago

@kinder_grinder There was nothing suspicious in the logs? Couldn't it be a dependency issue? The other day I had an issue at the restoring packages phase. I had a .Net application, and some nuget packages sources were unstable. Nothing I could do.