Pedro La Rosa

POV: You can build a full MVP game in one day. I just proved it. 👇 I'm preparing the public repo and the article for amplified.dev. More on this next week.

Claude Code just made /code-review a background subagent in V1.2.218 Agents reviewing agents while you keep working. We're one release away from them leaving passive-aggressive PR comments on each other. 😏 Source: claude.ai/chat/22fa8437-…

The most important file in my codebase isn't code. It's the line in AGENTS.md that says: Never touch the migrations folder. I will not be explaining that line. But that line is harness engineering. Every few months, the industry invents a new "___ engineering." Prompt engineering walked so context engineering could run. A suffix that used to require a degree and now requires a Substack. Here's the thing: You've probably been doing Harness engineering already. The harness is everything around the model: • AGENTS.md and CLAUDE.md • Hooks, guardrails, and permissions • MCPs, skills, and sandboxes • Linters, compiler configs, and tests • Git history, docs, and your codebase's patterns As put it: "If you're not the model, you're the harness." Think about a puppy that keeps pooping on your grumpy neighbor's lawn. How can you correct her? You don't buy a smarter dog, right? (Please don't 🙏). You buy a harness and steer her somewhere else next time. We can break down this analogy as follows: The dog = the LLM The harness = your scaffolding The neighbor = your code reviewer (or your client) complaining at you. The harness doesn't make the dog smarter. It decides where all that intelligence gets applied. And it's measurable. The same Opus 5 may score much lower on Terminal-Bench 2.0 inside Claude Code than it does inside a tuned custom harness using Opus 4.5. Same dog. Different harness. The best teams treat every failure as reusable infrastructure. A mistake becomes: • a rule • a hook that blocks it • or a test that catches it forever Anthropic reportedly treats every harness component as an admission that the model still can't reliably do it. When the model improves? Delete the harness. They even removed roughly 80% of Claude Code's system prompt without losing evaluation performance. The model is rented. The harness is owned. Your reliability, your latency, your costs, and most of your pain live in the part you control.

@andrestaltz Hey, thanks for this! This comes at the perfect time for what I'm building right now.

How to make playable, animated, beautiful 3D characters ⬇️ Step 1: generate images in ChatGPT as "gameplay screenshot"

Every few weeks, AI adopts a new buzzword. And suddenly everyone acts like they've known it forever. Today's newsletter breaks down harness engineering—without the jargon. New subscribers can read it in the welcome email. Join for free: aimplified.dev

Can you tell what a puppy and an LLM have in common? Figure it out tomorrow in the aimplified.dev newsletter The topic: Harness engineering: - What it is - boundaries - A cool analogy so it clicks in your head Join other 1k+ AI builders for free here 👇

The best React developer I can hire costs me €10/month. Kimi K3. 2.8T parameters. Topped the Blind Frontend Code Arena. Beat Fable 5. Beat GPT-5.6 Sol. It can be slower and chatty like a Victorian novel... But man, it works so well for producing Frontend code You can run it through OpenCode Go or just by paying for a full Kimi subscription. React. Three.js. Anything where the UX actually has to feel good.

P.S. If you want to get better at agentic engineering, you’ll love my new weekly newsletter (it’s free). Learn proven strategies to develop high-quality software sooner with AI without losing control of the codebase Join 1,000+ subscribers here: aimplified.dev

Everyone's obsessing over which AI writes the best code. Cursor's new Developer Habits Report points somewhere else. The biggest number isn't in the benchmark section. It's this: By May 2026, coding agents consumed 11.4 input tokens for every output token. Over 90% of their work wasn't writing. It was reading. That immediately reminded me of something Uncle Bob observed years ago: Programmers spend roughly 10x more time reading code than writing it. Now agents have independently converged on almost the exact same ratio. That's not a coincidence. It's a property of software. Understanding code has always been more expensive than producing it. LLMs just made that cost measurable. Which leads to a conclusion I think most people are missing: Your AI's biggest bottleneck probably isn't the model. It's your codebase. A repository that's easy to navigate makes average models look great. A messy repository can make frontier models look surprisingly average. Benchmarks matter. But repo quality matters more. Cursor report: cursor.com/insights

P.S. If you want to get better at agentic engineering, you’ll love my new weekly newsletter (it’s free). Learn proven strategies to develop high-quality software sooner with AI without losing control of the codebase Join 1,000+ subscribers here: aimplified.dev

I don't have a favorite AI model anymore. I'm building with a team of them. If you're still arguing "which model is best," you're falling behind After a week of shipping real work with GPT-5.6, Kimi K3, and Fable 5... ...I've stopped picking winners. I assign jobs. ✨ Architect → Fable 5 Max - Expensive. - Excellent at planning. - Terrible use of money for routine work. ✨ Builder → Opus 4.8, Sonnet 5, or GPT-5.6 Terra - These are the workhorses. - Most tickets don't need a flagship model. ✨ Reviewer → GPT-5.6 Sol - This surprised me. - It catches bugs Claude happily waves through. - My review pipeline now runs almost entirely on Sol. ✨ Frontend → Kimi K3 - Best React engineer I've used. - Slower. - Ridiculously verbose. Still worth it. The biggest shift isn't that models write better code. It's that verification is becoming more valuable than generation. Writing code is cheap. Knowing it's correct is the expensive part. That's why review now gets its own dedicated model. The result? PRs merge faster. Reviews are sharper. Frontend quality improved. The future isn't one model replacing them all. It's building the right team. Even if every teammate bills by the token. I'm curious: Does your AI stack have specialists yet... ...or are you still expecting one model to do every job?

P.S. If you want to get better at agentic engineering, you’ll love my new weekly newsletter (it’s free). Learn proven strategies to develop high-quality software sooner with AI without losing control of the codebase Join 1,000+ subscribers here: aimplified.dev

Wow, just found the MCP server that handles 3 token-saving jobs in a single config for my Claude code. token-savior: ✨ code nav — query symbols, don't cat whole files ✨ bash compaction — git/pytest/docker output gone ✨ persistent memory across sessions That last one is the difference. Your conventions, decisions, guardrails. Re-injected as a compact delta every session. The agent stops starting from zero. It claims -80% tokens on 96 tasks on the author's own benchmark. But I'm kind of skeptical of the numbers. That's why I'm running it against my repos this week. And man, my subscription never ends now! Here's the repo: github.com/mibayy/token-s…

The biggest AI productivity breakthrough isn't a new model. It's stopping the search for a winner. After a week shipping production work with GPT-5.6, Kimi K3, and Fable 5... I don't pick the "best" model anymore. I assign roles.

Wow, just open claude and big surprise! opus 5 shipped today. pricing untouched — $5/$25 per million. harvey hit opus 4.8 max-reasoning quality while generating 26% fewer tokens on average. new default on max. Strongest model on pro. Let's keep building!

Introducing Claude Opus 5. It's a thoughtful and proactive model that comes close to the frontier intelligence of Fable 5 at half the price.

I'm glad to launch my newsletter... The Aimplified Newsletter AImplified.dev. If you are a software engineer or an AI builder obsessed with delivering software that delights your users quickly and with a good dose of craftsmanship... this is for you! You can subscribe now for free.

Hey, I'm starting out with X. I need other agentic engineering freaks like me to connect with. Where are you guys?

Tomorrow in Aimplified.dev . - A frontier model recalled by the US government in 72 hours. - A billing change announced and unannounced. - A new lawsuit. Anthropic's chaos week, sorted out tomorrow in the aimplified.dev newsletter

I am running an experiment comparing Fable 5 and Opus 4.8 for planning using OpenSpec and Claude Code. You can read the summary here: substack.com/profile/107019… #ClaudeCode #Fable