Antonio Leiva

I'm Antonio Leiva. Turning AI into your ally as a developer. 🤖
Creator of @devexpert_io

Write about #AI, #AndroidDev, #Kotlin, & solopreneurship. 👇

[Antonio Leiva @leiva_dev]

@hamen How! Any suggestions? I have in my other account more than 10k followers, but maybe only 200-300 are verified

[Antonio Leiva @leiva_dev]

I'm on the street with my phone. I tell my @openclaw : "launch my app locally and create a URL to view it from my phone".

1 minute later I have a Cloudflare tunnel with the project running and accessible from the mobile browser.

Without opening the laptop. Without VPN. Without deployment.

What's happening underneath:
1️⃣ The agent executes `pnpm dev` on my machine
2️⃣ Opens a tunnel with `cloudflared`
3️⃣ Returns the public and ephemeral URL, which will stop working when I kill the process

How to use it?

1️⃣ brew install cloudflared
2️⃣ cloudflared tunnel --url http://127.0.0.1:3000

It's that easy.

Free, no logins or registrations.

This helps me keep moving forward without having to be sitting at my desk. This year I'm going to get a lot of sun 😄

Did you know about this? Do you think it could be helpful in your day-to-day?

[Antonio Leiva @leiva_dev]

@VasiliyZukanov The automations are in the Codex App.

I’m not using openclaw, just a much simpler tailor-made chatbot. So maybe you can do something similar (or even more powerful) with openclaw.

Even without the Codex App, this could be a set of cron tasks using agent CLIs in headless mode.

[Antonio Leiva @leiva_dev]

Days 4-5: Automations are autonomously increasing software quality.

Several expert agents detect problems, vulnerabilities, improvements, and another agent is launched periodically to implement the most prioritized issues.

Every day at 10 AM, my chatbot sends me a summary of the work done, and the issues that are unprioritized, with a priority recommendation based on relevance.

After my feedback, it prioritizes them and the rest proceed.

In a few days, if the prioritization is well-tuned, another agent will be in charge of prioritizing autonomously.

⚠️ Disclaimer: this is an experiment, under no circumstances am I recommending doing this in production environments. At the very least, I would have everything go through PRs (which I am not doing in my pet project)

[Antonio Leiva @leiva_dev]

Day 3: Now I have an army of experts programming

I had been wanting to try having agents work without my constant intervention for a while.

My first attempt was to create an automation in Codex App that would check the code every 2 hours and decide what to do. But it was too vague.

So, after talking with my aipal (my Telegram bot connected to Claude), we decided on this flow:

1️⃣ Every 2 hours, an "expert" on a specific topic reviews the code and creates issues on the most relevant topics they detect.

2️⃣ I manually review these issues and assign them 3 priority levels, or close them directly.

3️⃣ I can also manually create issues and assign them a priority.

4️⃣ The programming agent takes the oldest, highest-priority issue every X amount of time and implements it.

Again, I am the bottleneck deciding what gets implemented. I will iterate on that.

I'll tell you how it went in a future update.

[Antonio Leiva @leiva_dev]

A few months ago, this was unthinkable.

Yesterday, at the end of the day, I wondered: What if I could create the academy I've been dreaming of with a single prompt?

The idea was aggressive: I spent an hour answering questions with Codex's plan mode, outlining everything I wanted to see in it and how.

You've probably read that people have sessions lasting hours and end up with a complete app.

Well, the result wasn't quite like that, but it was the seed of something.

It generated a complete structure for my idea, which I can now iterate on with more precision.

The first thing I did this morning was sit in front of Pencil and start asking it for designs based on prompts that Codex generated.

The result? Very good, and after a few iterations, very close to how I had imagined it.

Then, using the MCP, Codex took it and updated the existing design with the new one.

I won't lie to you, it wasn't a one-prompt job, but we're starting to understand each other, and above all, we iterate on everything that fails so that the AGENTS.[md] is updated and doesn't happen again.

Right now, a day later, I have the dashboard page fully functional and, yes, with clearer ideas, I've left it working on the courses section.

Definition of Done: I can register a complete course with all its content. Will I achieve it in one prompt? I'll tell you tomorrow.

I still don't see how these tools can be given a perfect prompt to generate an app from start to finish, but for me, it was unthinkable to do this in a day while I've accomplished so many other things that had nothing to do with it.

The future academy is coming!

[Antonio Leiva @leiva_dev]

@hamen @openclaw Why? 😱

[Antonio Leiva @leiva_dev]

The future (or present) of software development is this.

The linear process is over, in a world where our attention is not focused on developing code but on orchestrating agents, we need to activate parallel thinking.

Creating specs while developing two features in parallel at the same time as fixing a bug is no longer a distant possibility.

If you don't do this, you are wasting the potential that AI offers you.

We know this well, and that's why in this edition of AI Expert we have transformed part of the curriculum to teach how to do this with basic tools, in any agent, and understanding how it works underneath.

If you're up for joining, we are already filling spots for the May edition 👇👇

https://t.co/yWV2ZVbut6

[Antonio Leiva @leiva_dev]

Yesterday was a BIG day in AI: Anthropic released Claude Opus 4.6 and OpenAI presented GPT-5.3-Codex.

Here's a technical breakdown of what each brings:

◼️ CLAUDE OPUS 4.6

Context: 1M tokens (beta), 128K output
Price: No change — $5/$25 per million (input/output)

Performance:
• +190 Elo over Opus 4.5 on economically relevant tasks
• Outperforms GPT-5.2 by ~144 Elo on GDPval-AA
• Higher score on Terminal-Bench 2.0 (agentic coding) and Humanity's Last Exam
• 76% accuracy on MRCR v2 (needle-in-haystack) vs 18.5% for Sonnet 4.5

New Features:
• Agent teams in Claude Code (autonomous agents in parallel)
• Adaptive thinking: reasons more when needed
• Effort controls (low/medium/high/max)
• Claude in PowerPoint and Excel improvements
• Lower over-refusal and new cybersecurity probes

◼️ GPT-5.3-CODEX

Unifies coding (5.2-Codex) + reasoning (5.2) into a single model, 25% faster.

Milestone: OpenAI's first model that helped build itself. Early versions were used to debug its training and manage deployment.

Benchmarks:
• Record on SWE-Bench Pro and Terminal-Bench 2.0
• Strong on OSWorld and GDPval
• With fewer tokens than previous models
• Real-time interaction without losing context
• Computer use far superior to previous GPTs
• Price not yet confirmed

Summary:
Opus 4.6 → Solid evolution. Better long context, reasoning, and agentic tools. Maintains prices.
GPT-5.3-Codex → Consolidation. Faster, more efficient, with the milestone of self-improvement during training.

The competition remains brutal.

[Antonio Leiva @leiva_dev]

🔍 Cursor, Cloudflare, Vercel, Cognition (the Devin folks) and Google just launched Agent Trace.

It's an open standard for answering a question more and more people are asking: who wrote this code, a human or an AI?

Agent Trace records metadata in each commit:
→ Which files and lines the AI touched
→ Which model generated it
→ A link to the original conversation

⚙️ How does it work?

The coding tool (Cursor, Claude Code, etc.) needs to implement the standard. Every time you accept generated code, the tool notes which lines the AI wrote, with which model, and in which session.

That metadata gets linked to the commit. It's not magic: if it's not active, there's no record.

What's it for?

📊 Auditing: knowing what percentage of your codebase comes from agents.
✅ Compliance: there are companies that need (or will need) this traceability.
🔎 Code review: having context of "this came from this conversation with Claude".

The interesting part is who's behind it. That Cursor, Cognition, and Google sat down together to define a standard says a lot about where this is heading.

The AI vs human authorship debate is no longer philosophical. It now has a technical specification.

👉 https://t.co/Ggprxmixch

[Antonio Leiva @leiva_dev]

The future of AI isn't a monolithic product; it's an open ecosystem of Skills.

Don't wait for someone to build the perfect bot; build the skills you actually need.

I’ve documented my full inventory here:
https://t.co/p11Qbps5Bg

[Antonio Leiva @leiva_dev]

Everyone is talking about "magic" AI agents like Moltbot (fka Clawdbot).

They promise to automate everything via WhatsApp, Calendar, and Web. It feels like magic, but the real superpower isn't the bot.

It’s the 𝗔𝗴𝗲𝗻𝘁 𝗦𝗸𝗶𝗹𝗹𝘀. 🧵

[Antonio Leiva @leiva_dev]

In my daily workflow, I’ve moved from "chatting with an AI" to having a team of specialists collaborating:

✅ Creator Synergy: Transcribe video -> Generate copy -> Schedule social -> Update newsletter.
✅ Admin Synergy: Detect invoice -> Extract data -> Upload to accounting.

[Antonio Leiva @leiva_dev]

@hamen @openclaw It’s not so heavy tbh, I think that with 32GB of RAM, it would be ok

[Antonio Leiva @leiva_dev]

My video about ClaudeBot / @openclaw is now dubbed.

It didn't went too bad!

The voice is not extremely similar to mine. I used Chatterbox-TTS.

Qwen3-TTS did a better job, but it was too slow on my 16GB MB Pro. Definitely a much better option if you can run it.

Lipsync would be nice, but I didn't try it, as I just wanted the audio for YouTube, and it doesn't allow dubbed video (only the audio)

You can check the result here, and learn a few things about how to use agentic chatbots:

[Antonio Leiva @leiva_dev]

@hamen @openclaw Yeah, not my accent, but definitely an accent 🤣

As it uses a recorded audio of myself talking in English, it tries to mimick that.

Qwen3-TTS was much closer to the reality. It's a pity I can't use it.

[Antonio Leiva @leiva_dev]

ClawdBot is awesome, it can do literally anything: read and answer emails, manage your calendar, write code, do PRs, even releases...

But there are a lot of concerns regarding security.

So I wrote this security guide to help you decide how much control you want to provide to the lobster 😁

https://t.co/w1raaAAr2v

[Antonio Leiva @leiva_dev]

@thethmuu_ @thekitze @steipete @MiniMax_AI But according to the video, it should

[Antonio Leiva @leiva_dev]

@thethmuu_ @thekitze @steipete @MiniMax_AI Not sure, better check this:

https://t.co/etLn7Er2bK

[MiniMax (official) @MiniMax_AI]

🦞To run a good Clawdbot, you need a model that has Trader Joes quality, Amazon Prime speed, and Temu prices.

🦞Here's how to get Clawdbot up and running with ClawdBotFather @steipete's personally recommended model, ✨M2.1: