Agents Claw, Code Factories, and 2026 Models

Key Takeaways

1"Claw" agents are becoming widespread, causing shortages of Mac mini and the emergence of agent social networks

2StrongDM implements software production rules without human writing or proofreading, inspiring other professionals

3The Claude Opus 4.5, GPT-5.1, and Gemini 3.1 Pro models enhance the reliability of coding agents

4Developers report a "Deep Blue" fatigue and a year of rapid upheaval

Why it matters : 2026 marks a concrete acceleration of agent-driven software automation, with tangible effects on developer practices, tools, and culture.

Between home-installed "Claw" agents, production rules without code review, and new, more reliable models, 2026 is shaking up software development. Concrete signals are emerging, from hard-to-find Mac minis to social networks for agents, while developers describe a year of transition and fatigue in the face of AI.

Visible Breaks: Rare Mac minis, MoltBook Boom, 100,000 Commits

In the Bay Area, Apple Stores have run out of Mac minis, purchased to run OpenClaw. Drew Breunig compared the purchase to an aquarium for a digital pet. Meanwhile, MoltBook, a social network for AI agents, launched on a Thursday, exploded on Friday, was featured by the New York Times on Monday, and then collapsed by Tuesday, overwhelmed by unwanted content and spam, before being acquired by Facebook/Meta a month later. On the software cadence front, the OpenClaw repository has now surpassed 100,000 commits. And Google, through Jeff Dean, released a video comparing Gemini 3 Pro and 3.1 Pro featuring animated animals on various modes of transportation.

Software Production Without Human Review: The StrongDM Method

In February, StrongDM introduced its "Software Factory," described by Dan Shapiro as a "Dark Factory." Since last July, the company has implemented two rules: code is not written by humans, and all code goes through a coding agent; human review is also prohibited. Once considered radical, these practices are now being adopted by some professionals, while code review and ways to bypass it dominated many sessions this year. StrongDM, a security company with decades of experience, is exploring how to verify quality without reading code, relying on agents. On the same front, about 40 of the 277 sessions at a conference focused on sandboxing or agent security.

From "Claws" to Personal Agents: A New Category Emerges

Born from an open repository launched in November and successively renamed CLAWDIS, CLAWDBOT, Moltbot, and finally OpenClaw at the end of January, the project surpassed 8,300 commits in less than two months. Described as very "vibe-coded," it catalyzed an "OpenClaw revolution" that helped define a new software category. The generic term "Claw" has emerged, with OpenClaw, NanoClaw, IronClaw, and PicoClaw, before the whole set was rebranded as "personal agents" or "general agents." In the San Francisco area, demand even emptied the stocks of Mac minis to host these agents.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Models and Coding Agents: A Milestone from November to February

Released in November, Claude Opus 4.5 and GPT-5.1 brought incremental gains, but their association with coding agents made their daily use more reliable. Claude Code has been around since February 2025, and Codex is more recent. A recurring test—requesting an SVG of a pelican on a bicycle—still showed weaknesses in November, including a poor bicycle frame for GPT-5.1. During the December holidays, many developers experimented with the model+agent tandem, and by January, enthusiasm was palpable to do more. In February, Google launched Gemini 3.1 Pro, capable of producing a coherent pelican on a bicycle down to the details, and Jeff Dean shared a video comparing Gemini 3 Pro and 3.1 Pro, teeming with animated animals on various transports, rendering some backup tests obsolete in light of the capacity increase.

Developers' Experiences: "Deep Blue," AI Mania, and DIY Projects

The term "Deep Blue," credited to Adam Leventhal, has been used to describe a fatigue among engineers in the face of AI, a theme widely echoed at conferences. The year has been felt as the fastest and most disruptive of their careers, with a need to understand its professional implications. In January, an episode of "AI mania" led to the start of overly ambitious projects, including a JavaScript interpreter entirely in Python, ported from MicroQuickJS by Fabrice Bellard, and a WebAssembly runtime in Python. A public demo showed the interpreter running via Pyodide—Python compiled to WebAssembly—executed in the browser. These experiences also served as a safeguard regarding their perceived utility while sustaining strong interest in WebAssembly. These observations and predictions were shared notably during the Oxide and Friends podcast with Bryan Cantrill and Adam Leventhal.

Predictions Met, Risks Avoided, and Kākāpō Season

Among the anticipations, the idea that LLMs would write good code is now considered achieved. Sandboxing has focused attention—with dozens of dedicated sessions—and the major accident feared for coding agents did not materialize. A joke mentioned a speech by the Pope on the economy of LLMs. More unusually, a prediction concerned an exceptional breeding season for Kākāpō: these nocturnal, flightless parrots numbered 236 at the beginning of the year, and the first chick in four years hatched on Valentine's Day, in a well-advanced season aided by a fruitful Rimu. This panorama fits into a 2026 still ongoing, which began as early as November 2025.