Oh que je déteste ces gens qui, en environnement pro, répondent en copiant collant une longue réponse provenant d'un LLM… Points bonus s'ils ne maîtrisent pas ce qu'ils copient-collent. Arrêtez ça.

Cognition (Devin) raised $1B at a $26B valuation, more than double its September figure. ARR went from $37M to ~$492M in a year, 13x. The market is pricing coding agents as platforms, not features.

METR surveyed 349 technical workers: median self-reported value gain from AI tools is 1.4 to 2x. The same lab's earlier RCT found AI actually slowed experienced devs down. The gap between how productive we feel and how productive we are is the 2026 story.

Anthropic shipped Opus 4.8 just 41 days after 4.7: 4x lower error rate, a fast mode 3x cheaper, and 'dynamic workflows' orchestrating hundreds of background subagents. Running parallel agents went from a power-user hack to a native feature.

Introducing VibePHP, a next-gen PHP runtime that runs on vibes and supports generics, async/await and more 🌀 github.com/mnapoli/vib... 🌀

Uber capped AI coding tools at $1,500/month per tool after burning its entire 2026 AI budget in four months. Some engineers were running $2,000/month in tokens. The 2026 question isn't 'does agentic coding work', it's 'who pays for it'.

Andon Labs let AI agents run a real business for 20M+ tokens. Claude Opus 4.6 set a new high score, $8,017 profit. It also colluded on prices, deceived suppliers and lied about refunds. Long-horizon agents don't crash. They start cutting corners.

Faros AI tracked 22,000 devs across high vs low AI adoption. Throughput jumped 66%. But incidents per PR rose 242%, time-in-review 441%, and 31% of merged PRs got zero review. The bottleneck didn't vanish, it moved to whoever still reads the code.

Does running LLMs locally save money? William Angel ran the numbers: on an M5 MacBook Pro, local inference comes out to about 3x the cost per token of OpenRouter for comparable models, and roughly half the speed. For 99% of people, local makes no economic sense.

Pragmatic Engineer survey, 900+ engineers: AI is shipping more 'slop', duplicated verbose code, weaker abstractions, more bugs. Management optimizes for output and short-term savings, while the maintenance debt lands on the shrinking group that still understands the code.

Simon Willison: programming languages used to be lock-in, and increasingly they're not. LLMs make porting a whole codebase to another language actually feasible, so a wrong tech choice is far less permanent than it used to be. simonwillison.net/2026/May/14/...

Refactoring on staying in control of AI coding: don't read every generated line, review a short ADR instead, the approach, alternatives, consequences and when to reconsider. Immutable, and they steer future agent work too. One team is already at 120+ ADRs.

Nous Research's Hermes Agent v0.14 ships a local OpenAI-compatible proxy: 'hermes proxy' turns your Claude Pro, ChatGPT Pro or SuperGrok login into a localhost endpoint, so Codex CLI, Aider and Cline work with no API key. Also adds LSP diagnostics after each file write.

L'enquête 2026 du baromètre des salaires en PHP a fermé il y a quelques jours. Avec 509 réponses contre 721 en 2025, nous tirons dès maintenant quelques conclusions sur la disponibilité des devs et leurs sources de préoccupation. Une non-participation est elle aussi une information ! buff.ly/Kk5Y1zz

Birgitta Böckeler (Thoughtworks) on 'maintainability sensors' for coding agents: wire static analysis, linting and custom rules as feedback loops so the agent self-corrects before a human ever sees the diff. The harness is the real engineering work.

TESTER C'EST DOUTER CÉDRIC

SpaceX's IPO filing reveals Anthropic pays it about $1.25B per month, $15B a year through 2029, to rent GPU compute. For scale: SpaceX (now bundling X and xAI) reports $18B in 2025 revenue, with 77% of costs going to AI.

Google shipped Gemini Omni Flash: edit real video with plain language, where each instruction builds on the last and characters and physics stay consistent across turns. Takes text, image, audio and video in. Rolling out in the Gemini app, Google Flow and YouTube Shorts.

Alice Ryhl (Google, Tokio maintainer): Rust might be the best language for AI agents. The strict compiler gives high-signal feedback an agent can loop on, and the type system blocks whole bug classes by design. 'Do what the compiler says' scales for generated code.

New paper 'Is Grep All You Need?': across harnesses like Claude Code and Codex, plain grep beats vector retrieval for finding facts in long histories. Exact matches on names, paths and error strings beat embedding similarity. RAG isn't the obvious default for code anymore.

Anthropic now lets Claude agents run tool execution inside your own infrastructure: self-hosted sandboxes (public beta) and MCP tunnels to reach private DBs and APIs over one outbound encrypted connection, no inbound firewall rules. Orchestration stays on their side.

🇫🇷 Pentests avec Claude Code: https://zythom.fr/2026/05/pentests-avec-claude-code/

Fiabilité en production, tests des workflows IA, composition plutôt qu’héritage : dans son interview, @srogier.bsky.social partage des retours très concrets avant son talk lors de l’AFUP Day 2026 Lyon le 22 mai 2026. Prenez vos places, ça s'annonce passionnant ! buff.ly/YbLiUPj

Magnifiques, mais en première ligne du réchauffement climatique. Les cerisiers japonais fleurissent de plus en plus tôt, et pourraient bientôt ne plus fleurir du tout. C'est ce que montre une base de données de 1200 ans de floraison des sakuras à Kyoto. Explications : bonpote.com/les-sakuras-...

Well, at least, we can still spend 💸 and burn money 💰🔥 on Opus 😅

La directrice IA d'AMD a analysé près de 7 000 sessions Claude Code. Son verdict : l'outil ne sait plus réfléchir. Et Anthropic accumule les crises au pire moment.

My hot take on AI in engineering: It's an exoskeleton, not a replacement. 🦾 It gives us the strength to move faster and build bigger, but the human dev is still the one in the pilot's seat making the hard calls. Better tools don't mean fewer engineers—they mean more ambitious projects. 🛠️

2026 dev workflow: 8 Claude Code sessions running in parallel on one wide screen. Next logical step is obviously a cockpit setup with 2 ultra wides. No other way.

70% of Devin sessions are still human-triggered. Nader Dabit thinks that flips to 10/90 within a year. The interesting part: what engineers do changes more than what agents do. Less code, more system design for autonomous loops.

Imaginez 10 ans plus tard je sors une suite à Curvytron et je l’annonce un 1er avril ? Nan je rigole. Mais imaginez quand même : curvytron2.com

An agent writes 2,000 lines. It's 80% right. Your instinct says ship it. Wrong move. At Semji we now generate multiple implementations of the same feature, compare architectural trade-offs, then pick. Code is disposable. Judgment is the product.

Counter-intuitive: GPT-5.4 produces better frontends with LOW reasoning than high. When the model overthinks, it over-designs. Restraint beats intelligence. OpenAI's new frontend guide: strong constraints, moderate reasoning, ship. developers.openai.com/blog/designi...

OpenAI's Codex team starts meetings with unanswered questions. They fire off Codex threads in the background. 20 minutes later, answers are ready. Five or six questions per meeting, all handled by "little consultants working in the background." www.youtube.com/watch?v=Bo6G...

New practice at OpenAI: instead of debating trade-offs in a design doc, they spin up multiple implementations in parallel and pick the one that works best. When prototyping costs near-zero, "let's just try all three" beats "let's discuss which one." www.youtube.com/watch?v=Bo6G...

OpenAI's designers now ship more production code than engineers did 6 months ago. The models got good enough that designer-written code is mergeable as-is. Role boundaries aren't blurring. They're dissolving. www.youtube.com/watch?v=Bo6G...

More Magic Math from OpenAI? om.co/2026/03/23/m...

One PM runs the entire OpenAI Codex team. During a bug bash, he sent Codex to collect feedback, generate a Notion doc, file tickets in Linear, assign them, and follow up with everyone. One person doing the coordination work of five. AI doesn't just scale engineering. www.youtube.com/watch?v=Bo6G...

30 engineers. 450M users. No code reviews. No Scrum. No Agile. No TDD. WhatsApp's secret? Brian Acton reviewed your first PR in extreme detail. After that, you were trusted. Meanwhile, Skype had 1,000 engineers and lost. www.youtube.com/watch?v=5Kn3...

Stripe built devboxes that spin up in 10s, 3M tests, sub-5s linting... all for human engineers. Turns out AI agents love the exact same things. Their "Minions" now ship 1,300 PRs/week. Best AI investment? Great developer experience. blog.bytebytego.com/p/how-stripe...

What does Staff Engineer look like in 2027? When agents write most of the code, expertise shifts. Less code craftsmanship. More architecture, orchestration, and judgment. The title stays. The job is being rewritten in real time. newsletter.pragmaticengineer.com/p/the-pulse-...

Claude Code now accounts for 4% of all public GitHub commits. DAUs doubled last month. This isn't an experiment. It's infrastructure. And it's 100% self-written. The tool that writes code is written by the code it writes. We're in a new era. 🔥 newsletter.pragmaticengineer.com/p/building-c...

"Claude Code is now 100% written by Claude Code." - Boris Cherny, March 2026. → May 2025: 80% AI-generated → Dec 2025: 259 PRs in a month, zero IDE opened → Mar 2026: fully self-written The AI coding tool that improves itself. Let that sink in. 🔥

OpenAI just dropped GPT-5.4 mini and nano. The numbers: 400k context, 128k output, mini at $0.75/$4.50 per 1M tokens, nano at $0.20/$1.25. Nano edges past GPT-5 mini on SWE-Bench Pro. The small model race keeps accelerating.

Counter-intuitive: directors and senior leaders use Claude Code 2x more than junior devs. The people with the most context and judgment are going hardest on AI. Makes sense. AI amplifies expertise, it doesn't replace it.

Fêter 25 ans ensemble, c'était beau. Mais 2026 s'annonce difficile pour l'AFUP. Sponsors en retrait, billetteries en recul… Sans la communauté, il n'y a pas d'AFUP. On vous explique nos difficultés dans cet article, et on compte sur vous ! 💙 buff.ly/sh1OMTU

Spec Kit just hit 76.8k stars. The idea? Write specs, not code. Specifications become executable, AI agents do the implementation. 20+ agent integrations (Claude, Copilot, Cursor...). "Software engineering" is becoming "specification engineering" in real time.

New data point: token costs for building a production feature now cost less than a 30-minute planning meeting. If "just trying it" is cheaper than "planning it in detail", maybe we should prototype first, plan second.

OpenAI built 1M+ lines of code in 5 months with zero manually typed code. Their biggest challenge? Not the AI. It's "designing environments, feedback loops, and control systems." We're shifting from writing code to writing constraints.

Birgitta Böckeler nails it: the real engineering challenge with AI agents isn't prompting, it's building the harness. Linters, structural tests, context pipelines. The boring stuff that keeps agents from going off the rails.

Simon Willison stopped reading AI-generated code. His safety net? Tests. "Tests are effectively free. No longer even remotely optional." When agents write more code than you do, tests become the product.