Jose Luis Matus

CTO building with AI, leading 40 engineers remotely. Writing about what actually works →

[Jose Luis Matus @jmatux]

@torrenegra That’s a bold filter. Most CTOs I talk to have a similar stance but never pull the trigger because migration risk feels too high. What surprised you most once you actually committed to building the replacements?

[Jose Luis Matus @jmatux]

@nicowalter25 He pensado en esto y me preocupa que si hacemos agentes que revisen el código y tenemos éxito con ello quien se queda con el conocimiento del sistema ? Escribe algo relacionado basado en mi experiencia : https://t.co/w2grFs1bb1

[Jose Luis Matus @jmatux]

@plainionist I agree I think now we have a knowledge problem https://t.co/AHVkrTkNIp

[Jose Luis Matus @jmatux]

@yacineMTB Makes sense due to the output generated by AI. I believe code review is importante for knowledge , perhaps it needs to change not eliminated .

[Jose Luis Matus @jmatux]

What's helping on my teams:

→ Cross-dev testing: dev 1 tests what dev 2 built. Preserves system knowledge, adds the adversarial lens.
→ Agent-assisted automation for regression volume.
→ Keeping at least one person with comprehensive system exposure — and protecting their time.

Honest answer: it's not solved yet.

The question worth sitting with: who on your team can describe how your system actually behaves, end to end, from memory?

If you're not sure → that's the gap.

Full piece: https://t.co/ANXW8QFYfN

[Jose Luis Matus @jmatux]

Last month I reviewed a PR and realized I was the only human who'd read it.

Developer prompted an agent. Generated code. Quick test. PR opened. Approved in minutes.

Everyone knew what it was supposed to do. Nobody read every line that got there.

This is the quiet crisis in AI-accelerated teams. 🧵

[Jose Luis Matus @jmatux]

From CodeRabbit's analysis of ~500 open-source GitHub PRs:

AI-co-authored code has 1.7× more issues than human-only code.
Logic errors: 75% more common.
Security vulnerabilities: nearly 3× as frequent.

The testing surface is expanding.
The people covering it are losing ground.

[Jose Luis Matus @jmatux]

@cvander breaking down Valerie, an AI agent running a real vending machine in SF. The headline isn't the demo. It's the legal stack: giving an agent business ownership through a trust, and the regulatory gap that still surrounds it. Agentic commerce shipped to a real corner of SF.

[This Week in Startups @twistartups]

An AI agent named Valerie is running a real vending machine in San Francisco and it’s not (just) a stunt. The OpenClaw-powered vending machine is helping @cvander find the limits of agentic commerce present today — and what needs to be built next to empower our synthetic friends.

@Jason and @Alex also welcomed @0xcarro on the show to chat his company Manifold, its Targon compute product, and the Bittensor network that underpins it. Then it was time to parse Big Tech earnings, talk bounties and close out the week with a dose of the NBA.

0:00 Intro & sponsor reads (Pilot, Shopify, Grasshopper Bank)
1:06 Christian van der Henst: Valerie the AI vending machine demo
2:23 How OpenClaw powers Valerie: Dynamic pricing, inventory, bank account
4:28 Legal structure: Giving an AI agent business ownership via trust
7:23 Where agents can and can't operate today
8:46 How agentic business ownership works legally
9:46 Jason's SFO café, and SF regulations vs. AI-run businesses
10:10 Grasshopper Bank: Time is money. Don't waste either. Go to https://t.co/83Lr7qiYZU and get an exclusive $500 cash bonus just for opening an account.
11:48 AI café in Stockholm running on agents
18:21 Plaud: If your work depends on conversations — interviews, meetings, calls — you need a Plaud NotePin. You can check it out at https://t.co/AhhYi7C73V and use code TWIST for 10% off!
19:05 Robert Myers, Manifold Labs: Targon & Bittensor Subnet 4 interview
20:21 What is Bittensor? An "incubator with 128 subnets"
21:18 Shopify: Turn those What If's into sales with the ecommerce platform powering millions of businesses. Sign up for your $1-per-month trial today at https://t.co/JtDqtR5bLP
21:43 From Sybil to Targon: Manifold's journey to confidential compute
25:21 Pricing, utilization caps, and why GPUs are sold out
26:38 Who's using Targon? Customers, use cases, and the mom-and-pop data center argument
30:06 Pilot: Focus on your product, let Pilot handle your bookkeeping. Pilot provides the most reliable accounting, CFO, and tax services for startups and small businesses. Head to https://t.co/LEOcqgC0Vg and get $1,200 off your first year.
34:36 Bounties update: podcast companion (May 8) and https://t.co/HLCBbO6LKC (May 18)
35:07 Jason explains the https://t.co/jg4IxOq5KY vision — 15 years in the making
39:11 Polymarket: Will Anthropic flip Bitcoin by Dec 31?
40:29 Jason's Bitcoin bear case: "It's played out. No incremental buyers."
44:09 Alex's supporting points: Stablecoins are replacing BTC's killer use case
46:05 MicroStrategy / Strategy updates
47:48 Big Tech earnings: Google Cloud +63%, AWS best quarter in 15 quarters
52:37 AI compute demand vs. the fiber overbuild analogy
55:55 Congress pressuring startups over Chinese AI models (DeepSeek, Moonshot)
57:11 A16z on the geopolitical risk of Chinese AI models
1:00:08 Reflection AI — America's open source AI champion (or lack thereof)
1:01:26 Off Duty: Knicks blow out Atlanta Hawks 140–89, Jason goes road-tripping
1:03:06 Tip: See your team play in another city's arena
1:04:15 Sixers vs. Celtics preview; NBA Playoffs Polymarket odds

🎥 Watch the full episode here 👇

[Jose Luis Matus @jmatux]

Full report on the 240 hours, including the parts where the experiment strained:
https://t.co/zbkEQTIjsB

[Jose Luis Matus @jmatux]

I ran Codex and Claude Code side by side for 240 hours shipping a production mobile app.
Both shipped real code. They were not interchangeable.
Here is where each one actually beat the other.

[Jose Luis Matus @jmatux]

The split I settled into:

Codex for plan-heavy structural work. Service layers, test coverage, repetitive feature scaffolding.
Claude Code for UI work and debugging loops where I needed to paste screenshots back.

Both subs paid for themselves.