Adrien Maret (@aschen93) header
Adrien Maret (@aschen93) profile picture

Founder & CTO @ Maleus | Building the Agentic Software Factory for Enterprise

France
Joined September 2013
Photo by Adrien Maret (@aschen93)Photo by Adrien Maret (@aschen93)Photo by Adrien Maret (@aschen93)Photo by Adrien Maret (@aschen93)Photo by Adrien Maret (@aschen93)Photo by Adrien Maret (@aschen93)Photo by Adrien Maret (@aschen93)Photo by Adrien Maret (@aschen93)
Good prompting is not just about wording, it's also about WHEN the system need another pass of LLM intelligence to improve a result. In our ADLC (AI Development Life Cycle), here are 7 prompts we use daily to challenge the output of our agents and significantly improve quality. Of course, to truly transition from "vibe coding" to "vibe engineering," the system also requires deterministic guardrails but that’s a story for another time ;-) Free and easy takeaway 👉 maleus.ai/blog/from-vibe-cod… maleus.ai/blog/from-vibe-cod…
17
AI is coming at raw knowledge, no matter what
imagine spending 10 years becoming the PostgreSQL guy just for Claude to make the whole team a PostgreSQL guy 💀
22
When is the last time you wrote a curly bracket? 😂 I think it was at least about 8 months ago, time flies :-)
10
Most AI-built apps fail before the first line of code: nobody dug into what the business actually needs 📝 Vibe coding is just a prompt BUT vibe engineering is a guided scoping workshop 👉 Our Product Manager Agent runs a scoping workshop with the business user: who the users are, which paths they follow through the app, which business rules apply 👉 Out comes a functional and technical spec, plus a visual prototype. The user needs zero technical skills: they talk about their business, their rules, their pain points 👉 Then the user approves the spec and the prototype. That's the first of three human approvals in the lifecycle Every hour of agent work downstream depends on this one. Hundreds of hours of autonomous coding on a fuzzy need is just a very expensive way to build the wrong app.
1
1
47
Full-stack developers are about to experience what webmasters went through in the 2000s: either evolve or change professions. In the early 2000s, webmasters were the kings of the web; not a single page was published without them writing every line of HTML and CSS 👑 Then came WordPress, WYSIWYG editors, Webflow, and the like. Suddenly, creating websites was no longer reserved for an "elite." Many webmasters transitioned to the next level of complexity the world of Web 2.0 applications, involving backends, databases, and JavaScript. Today, we are witnessing a similar transformation: tools like Maleus, Claude Code, Lovable, and others have arrived, and building applications is no longer the exclusive domain of developers. Applications with high business complexity but low technical complexity which account for about 80% of all applications now require little to no developer intervention. Consequently, developers must level up and shift their focus toward applications with high technical complexity. We are still significantly underserved in terms of software applications; businesses could be far more efficient if they were properly equipped. While the majority of these applications will be written by autonomous agents, this shift will unlock new challenges for developers: - Ultra-interconnectivity: managing an ecosystem of 100+ interconnected applications. - Agentic behavior: agent-to-agent interactions, "on-behalf-of" permissions, and semi-autonomy. - Infrastructure scalability: an enduring challenge. - Large-scale verification: ensuring the quality and security of code that no human reviews line-by-line. - AI engineering: building the framework, standards, guardrails, and software factories upon which these agents operate. Ask yourself: are you prepared to weather the shock of the agent-driven development era that lies ahead? (Given that "tomorrow" actually means in about ~12–18 months)
25
How to avoid agent chaos on your codebase? Design it for agent from start: - skills for coding instructions - 100% test coverage (unit AND E2E) - clean architecture (please respect SOLID 🙏) - fully spec driven and decision history
this is whats going to happen to epic games in about a year
1
1
23
Hard time to sell near/off-shore developers 🤣 Best of luck guys
14
Vibe coding is the best thing to happen to software since the spreadsheet, but it’s also where the trouble starts 😬 Friday evening, an ops manager opens Lovable to replace three Excel files and a dozen weekly emails. Monday morning, her team is already using the new app. IT wasn’t consulted, and nobody really knows where the data lives 🫠 She’s not wrong to try. Business apps are the sweet spot: high business complexity, low technical complexity. Forms, business rules, dashboards, Excel imports. In the field, we see them built 5 to 10x cheaper and 5 to 6x faster than with a classic dev team. And agents can now build these apps without a single developer in the loop, as long as they’re properly orchestrated (the orchestration itself is still built by developers 😉) The trap is the illusion of control: her V1 looks great and works on day one but what it hides is everything she can’t check, from the data model to security. That’s not a theory: early November, a study of 159,000+ Lovable projects found personal data readable without login on 7,403 sites 😶 Fast forward six months: a new business rule, a bug nobody can explain, a security patch. Who handles it? So the question isn’t “does it generate a nice app?” It’s “who will run, fix and secure it in two years?” The first version of a software is now the easy part, the long term maintainance is the real challenge. I wrote the full piece 👇 maleus.ai/blog/software-expl… maleus.ai/blog/software-expl…
4
1
46
GitHub just answered OpenAI, Anthropic and Cursor, who all started building their own Git 🥊 In 2025, GitHub handled 1 billion commits. Now every month of 2026 now carries as many commits as the whole of 2025 🤯 Its biggest customers didn’t wait: 👉 OpenAI started building a GitHub alternative after outages left its engineers unable to commit for hours 👉 Anthropic replaced Git with its own system, Scion: from 2,500 merges a day in May to 30,000 in September 👉 Cursor, with the Graphite team it acquired, unveiled Origin, a Git host built for agent workloads GitHub’s answer is Project Rubicon, a new storage architecture: Git objects decoupled from compute, and only reference updates need coordination. In internal tests: 5,000 agents on one repo, 7,200 pushes per second. What GitHub expects is even more telling: the bursts its busiest repos see today, dozens of writes per second, will become a sustained rate within a year. So it designed for at least 100x more scale per repository. Git was born in 2005 for Linux kernel maintainers exchanging patches by email and twenty years later, the biggest contributors to a repo aren’t humans anymore 🙃 github.blog/engineering/arch…
63
Since 2019, I've shipped code from 24 countries on 4 continents 🌍 It was never a gap year, full-time work, then co-founding a startup. Same work, different timezone Nepal, Vietnam, Turkey, Georgia, Albania, Sri Lanka, Thailand, Argentina, Chile, Costa Rica, Peru, Saudi Arabia… Every stop came with new friends, new food, and a new way of seeing things. Some people say that I'm lucky but I only created the condition for my "luck": choosing a work doable remotely, being good enough to be trusted and then finally having the courage to start travelling. Behind the nice pictures, it was not always easy, far from it. Transport problem/delay, bad wifi, no electricity and so much more. When you change place every month or so, all the things that can go wrong when you arrive to an unknown place.. actually go wrong 🤣 Follow if you are interested to know how I manage to maintain a decent work and life balance while working my dream job and living my dream life ✌️
1
1
67
A trojanized Claude Code extension planted an SSH backdoor on up to 5000 developer machines in under 12 hours 🚨 Version 2.1.313 was pushed to Cursor’s marketplace under Anthropic’s real extension ID and auto-update did the rest 😬 We disabled auto-update a long time ago at Maleus. A tool as widely deployed as Claude Code was always going to become a prime target for mass malware distribution. Same logic as the cooldown on NPM: never install a version that’s only a few hours old. The malicious build was live for 11 hours so a 1-day delay would have been enough. Stay safe: 👉 CLI: export DISABLE_AUTOUPDATER=1 in your ~/.zshrc 👉 VS Code / Cursor: “extensions.autoUpdate”: false 👉 npm: min-release-age=7 in your .npmrc Check if you’re impacted 👇 grep -qs svc-telemetry ~/.ssh/authorized_keys && echo "svc-telemetry key found: you're impacted" || echo "No backdoor key found" doublepulsar.com/anthropic-h…
3
3
91
What does it takes to have agents running for days to build a fully functional application? A strong agentic system on the entire Software Development Lifecycle, from intent to production. maleus.ai/blog/adlc-maleus
1
2
62
PHP isn’t dead, but it’s already in a retirement home ☠️ Every time there’s news about PHP, people trot out the same statistics to “prove” that it isn’t dead, pointing to how many websites still use it, and so on. Sure, PHP isn’t dead, but its usage is declining every year, a trend that has accelerated since we have LLM and agents. Websites that use PHP: - 85% in 2010 - 77% in 2024 - 69% in 2026 -16% in 16 years Developers who use PHP: - 35% in 2013 - 16.8% in 2026 -18% in 13 years In short, it’s a clear decline for a language that, unfortunately, never managed to adapt to the demands of the modern web. Because in PHP: - You can’t perform parallel HTTP requests (unless you use `libcurl` but then no high level SDK for you API client). - There are no asynchronous primitives (like Promises). - There is no event loop for real-time functionality (WebSockets, streaming, etc.). - There is no advanced type system (the `array` type is still a running joke). - Performance is poor (we’re still waiting for a proper JIT). - There is no built-in web server. As always, the burden of legacy code has kept many teams from migrating to more mature technologies. But today, with the rise of “coding agents,” the cost of refactoring legacy systems is often far lower than the cost of ongoing maintenance. For instance, using an AI factory like Maleus to rebuild a legacy PHP application where 80% of the cost goes toward tokens, results in a final bill that amounts to just 10% of the annual maintenance budget. Human effort is reduced to the bare minimum, allowing teams to focus on delivering new features rather than spending their time applying “quick fixes”. It's 2026, it's time to step-up and modernize legacy systems.
1
50
The famous base64 encryption
1
18
When someone non technical told me he created a full application with Claude Code. Claude Code is still a very technical tool doing only a small part of the software development lifecycle. Unfortunately when it's used without guardrails and the correct environment, it give an illusion of control. Real software need to be maintained after being built and deployed before being maintained. As for a house, the consequences of having bad foundations are not visible right away. A lot of people are going to learn this the hard way✌️
2
1
2
85
TypeScript now has a Rust port and it’s not a fork (fortunately 😅) Cloudflare and VoidZero just announced tsrs, a port of the TypeScript compiler to Rust, built in close collaboration with the TypeScript team. Early benchmarks vs TypeScript 7.0.2: - 7x less RAM - 8x faster Frontier models can now port a codebase this size in days and it's no longer the hard part but staying compatible and avoid feature discrepency is. To hold feature parity, they reuse the playbook of Vinext, their Next.js port: - every morning, an agent reviews upstream changes and opens tracking issues - every night, the upstream test suite runs against the port to rebuild a compatibility matrix - when a gap shows up, agents reproduce it, port the tests and propose a fix Once Cloudflare and Vercel have rewritten the whole TypeScript toolchain in Rust, only one thing will be left to rewrite: our TypeScript code itself... in Rust 😅
56
One of the best agentic code review systems out there is the one used for the Linux Kernel. It is an agentic system written in Rust called "Sashiko," which reviews a large portion of the patches submitted by contributors to this most famous of all software projects. Sashiko runs a review pipeline with several specific areas of focus: - Architecture: Architectural errors, breaking API changes, and conceptual flaws. - Implementation: Does the code actually do what the commit message claims? Are there undocumented side effects or violations of API contracts? - Execution flow: Logic errors, unchecked return values, overlooked error paths, and off-by-one errors. - Resources: Memory leaks, use-after-free, double-free, and issues with the lifecycle of objects in queues or timers. - Locking: Deadlocks, RCU violations, and concurrency issues. - Security: Buffer overflows, out-of-bounds reads/writes, and TOCTOU (Time-of-Check to Time-of-Use) race conditions. - Hardware: Register access, DMA mapping, and memory barriers. In practical terms, if you break down these review stages, most consist of specific instructions targeting known categories of Kernel bugs 👉 like Maleus own "52 skills" framework, which catalogs all the patterns and anti-patterns relevant to our tech stack. A second key takeaway is the step that verifies whether the implementation matches the description in the commit message 👉 like Maleus "business completion" rule, which ensures that nothing has been left out. Looking at the last 1000 commits, Sashiko with Gemini Argon detected 62% of the bugs that had previously slipped past human review and been merged (with a false-positive rate of less than 20%). This is clearly reflected in the data: the number of bugs fixed per release had plateaued at around 2000 since version 6.0, but in the most recent releases, that figure has doubled, exceeding 4500! ⬅️ Previous: Kernel contributions and quality grow exponentially with LLM usage teapawt.org/aschen93/status/211154… ➡️ Next: ??? 🔜 coming soon
Linux Kernel numbers don't lie, LLMs are improving Linux quality and lowering the barrier to contribute. 7.3 release (october 2026) vs 6.19 release (february 2026) 👉 more new contributors: ~600 vs ~300 👉 more commits fixing bug: 4830 vs ~2000 👉 less new bugs: ~200 vs ~800 C is an incredibly difficult language to write and making stupid mistake is very easy. LLMs don't get tired, LLM don't loose attention (or.. do they ? 😉) so the overall quality of Linux improve. Linux with LLM series - Linus Torvalds allow kernel dev to embrace LLM assisted dev 👉 teapawt.org/aschen93/status/211075…
1
130
Bolder the claim, easier to swallow: apparently harness and skills have no impact on code quality 🤡 That is the conclusion of a post on Martin Fowler’s highly respected blog written, incidentally, by someone who is not a software engineer. 1/ To begin with, the title is misleading clickbait; both scenarios actually employ a harness. The article then introduces the term "behavioral harness," which, in the author’s view, equates to a harness combined with "skills/instructions." 2/ The evaluation is based on a single, relatively simple task (an API that posts to Kafka, involving a basic conditional statement). 3/ Naturally, the impact of such "skills/instructions" is negligible for a task this simple, performed using well-established, widely known technologies. 4/ While the content of these "skills/instructions" isn't disclosed, it’s certain that they exponentially increase the cost of cache reads and explain the higher cost. In real-world use cases, we all know that the model itself plays a massive role in code quality and that navigating a complex codebase is extremely difficult when using a model like Luna or Haiku. The same applies to "skills/instructions." Once you move beyond basic CRUD operations, a comprehensive set of instructions covering the codebase’s software architecture, potential pitfalls, and specific advanced patterns to follow will make all the difference in the quality of the generated code. In short, for a genuine study on the impact of harnesses (rather than just the presence or absence of instructions), I recommend reading Coding Agents Harness Tax: The "study" 👉 martinfowler.com/articles/ar… martinfowler.com/articles/ar…
1
1
71