Pedro La Rosa

Leveraging AI to ship high-quality software faster that meets the requirements, and sharing everything I learn along the way @ ⚡

[Pedro La Rosa @PedroLaRosaDev]

You're not a software engineer anymore...

But still shipping productive software.

How could it be?

In this article, I'm breaking down the latest InfoQ Culture Trends 2026 and highlighting the most valuable takeaways on AI software development. https://t.co/PMwIoqEGKc

[Pedro La Rosa @PedroLaRosaDev]

Haha, It's so clever. 😂

[terminally onλine εngineer @tekbog]

code review in 2026

[Pedro La Rosa @PedroLaRosaDev]

Unpopular opinion:

Many respected devs have stopped reading code.

Not reviewing less.

Not skimming faster.

Stopped.

Dangerous when real users and revenue depend on what ships.

Tomorrow on AImplified: how to review agent diffs without line-by-line archaeology.

Link in bio

[Pedro La Rosa @PedroLaRosaDev]

Damn…

AI now writes 42% of committed code.

96% of devs don’t fully trust it.
Only 48% always check it.
We’re just getting started.

Be honest:

Did you think AI was already writing this much of the code we ship?

I got this insight from: https://t.co/uQvEttr7K4

[Pedro La Rosa @PedroLaRosaDev]

The 2026 InfoQ report renamed our job.

Engineers are becoming custodians, not contributors.

Not writing code. Steering agents, checking what comes back, wrapping it all in constraints that hold.

The quiet part, in print.

https://t.co/79eNprHjQz

[Pedro La Rosa @PedroLaRosaDev]

You're not a software engineer anymore...

But still shipping productive software.

How could it be?

In this article, I'm breaking down the latest InfoQ Culture Trends 2026 and highlighting the most valuable takeaways on AI software development. https://t.co/PMwIoqEGKc

[Pedro La Rosa @PedroLaRosaDev]

I've got a great kindle library with awesome SW classics like refactoring, clean code etc...

I read them. But sometimes just don't remember the important bits, or which book I read what in.

This happens to you too, right?

Well, last week I found an agent skill that solves this altogether.

book-to-skill, by virgiliojr94. it converts a book (PDF, EPUB, DOCX, Markdown, even MOBI via Caliber) into a proper agent skill.

you name it:

frameworks, decision rules anti-patterns, per-chapter files your agent loads on demand

Ask about any example, any fragment, and it answers from the actual chapter, instead of hallucinating the general aura of the book.

It works beyond books: fold a docs/ folder, your adrs, or a stack of papers into a single skill.

Now I have a smart personal librarian, always available to help me recall important things from my digital library 🤓

Link here: https://t.co/PNEh4nvSaj

[Pedro La Rosa @PedroLaRosaDev]

Code generation is the easy part.

Persistent context and a structured SDLC for agents feel like the real unlock.

Also, defining clear, deterministic guardrails (lints, tests, e2e tests, etc.) to keep the code maintainable and consistently aligned with your standards so it can scale sustainably over time

[Pedro La Rosa @PedroLaRosaDev]

@AnupamHaldkar It is so impressive how we went from no AI at all to auto-completion, and now to delegating the full code to AI.

All of this in less than 4 years 🤯.

So crazy, bro.

[Pedro La Rosa @PedroLaRosaDev]

Amen sir,

That's now what the difference is between a software engineer and anything else.

One thing is to build and release the app. That's the easy part. The complexity is maintaining and extending the software.

And man... If your software has some success. You will want to have your customers happy.

So, that's where a SW is so valuable and can take that app further.

[Pedro La Rosa @PedroLaRosaDev]

Hey, I appreciate the no-hype practical take.

Fully autonomous agents still feel more theater than productivity for a lot of real work.

However, There's a sad reality we can't ignore.

Code is getting less relevant; "good enough" is enough to go into production.

Now we need to define clear, deterministic guardrails (linters, tests, e2e tests, docs, etc.) so the Agent can automatically run quality checks to maintain a scalable, clean codebase, reducing the effort required to read thousands of LOC diffs line by line.

[Pedro La Rosa @PedroLaRosaDev]

@chaosengineerr Haha, man, that's so real.

The ideas explosion goes even faster than the number of opened sessions with Claude xD.

What I'm doing now is researching the idea and sketching up the initial MVP by hand and think about feasibility before sending the first prompt.

[Pedro La Rosa @PedroLaRosaDev]

Two-pizza teams are becoming two humans plus a swarm of agents.

faster? yes.

Also a recipe for lost diversity of thought and extra-crispy burnout.

Read more here: https://t.co/uQvEttqzUw

[Pedro La Rosa @PedroLaRosaDev]

Context rot is real.

The longer an agent session runs, the more stale assumptions pile up.

Sometimes the fastest move isn’t “keep going.”

It’s creating a handoff.md and starting a fresh session.

How often do you start a new session with your agent?

[Pedro La Rosa @PedroLaRosaDev]

Don’t trust benchmarks blindly.

Vendor numbers are useful.

But when a new open model claims to be the best coder, test it yourself.

Give it the same hard task, Minimal harness, and the same success criteria.

Your own codebase is the benchmark that matters.

[Pedro La Rosa @PedroLaRosaDev]

@mattshumer_ You can play here: https://t.co/HA4HfF7ix9

The game's still in beta. But it is totally playable.

Enjoy

[Pedro La Rosa @PedroLaRosaDev]

A few weeks ago, I created a new game using @mattshumer_ 's gauntlet loop.

An ultra-fast, antigravity racing game where you can drive 2000 km/h.

Yesterday, I released a new track:

Halcyon Reach, inspired by Avatar's Pandora's floating mountain.

Play in the first comment 👇

[Pedro La Rosa @PedroLaRosaDev]

@AnandButani Solid round-up, dude.

OpenCode has been my daily horse lately; it's awesome to have such a catalog of LLMs to use.

It is a pity you can't use your Claude subscription with it 🙁

[Pedro La Rosa @PedroLaRosaDev]

@suraj_sharma14 1 to 3 are the same problem in disguise (or at least related).

The lack of accountability and rigor is what's breaking apps.

What is delivered goes with your signature. The "AI did this" won't help when you have to explain to your manager/clients about that critical bug.

[Pedro La Rosa @PedroLaRosaDev]

Breaking: AI produces 42% of committed code according to InfoQ’s culture trends report 2026.

The scary part? 96% of devs don’t fully trust it, while only 48% review it.

We're breaking down the full report tomorrow in the https://t.co/zKbpeE9HYj newsletter.

See you there! 😉

[Pedro La Rosa @PedroLaRosaDev]

POV: A new model is only as good as the harness underneath it.

The same weights can feel incredible or mediocre depending on:

Tool quality.
Context management.
Retry logic.

Stop evaluating models in isolation.

Evaluate the entire system.