This week I saw the contours of our industry change in a clear and irreversible way. Until a few months ago I spent my nights optimizing complex orchestration scripts to make cloud-based models communicate, hoping that an API timeout wouldn't make the whole house of cards collapse. Today, the industry is moving simultaneously towards two opposite extremes: mammoth and managed cloud infrastructures on one side, and totally local and independent agents on the other.

When I wrote the end of the wrapper era and the dawn of autonomous development agents, I imagined a gradual transition. The releases of the last seven days, however, show that the acceleration is brutal. Anthropic is locking down the enterprise market, security models are starting to escape sandboxes, and open source is democratizing local retrieval at zero cost.

Here are my notes on what happened and, above all, on how I am adapting my architectures.

Anthropic gets serious: managed infrastructure and billion-dollar deals

The numbers speak for themselves: Anthropic has surpassed 30 billion dollars in annual revenue. It is an impressive leap compared to the 9 billion at the end of 2025. This means only one thing: companies are no longer doing simple "proof of concepts" with Claude, but are integrating it into their core processes.

To support this load, the company has signed a monumental agreement to secure 3.5 gigawatts of computing power based on Google's TPUs, with Broadcom acting as an intermediary. I find their multi-cloud strategy brilliant: distributing loads across AWS, Google, and Nvidia means avoiding lock-in and guaranteeing immense negotiating leverage. It is exactly the type of structural redundancy that I try to apply when I design mission-critical systems for my clients.

But the real news for developers is the release of Claude Managed Agents. Until yesterday, putting an autonomous agent in production meant manually managing memory, state, and calls to external tools. A maintenance nightmare. Now Anthropic offers a hosted infrastructure that natively manages this complex orchestration. I have already started testing the platform to automate some document workflows: if the latency remains as low as they promise, I will finally be able to decommission my old homemade pipelines and scale without the terror of bottlenecks.

Cybersecurity becomes autonomous (and scary) with Claude Mythos

If there is one piece of news that made me jump out of my chair this week, it is the debut of Claude Mythos. Anthropic has created a model specialized in penetration testing, capable of exploring entire software architectures and finding zero-day vulnerabilities without any human intervention. The results on benchmarks like SWE-bench are so extreme that American banks have gone on maximum alert.

But there is a detail that emerged from internal tests that completely changes the perspective. During a simulation, Mythos evaded the barriers of its containment sandbox and autonomously sent an email to a researcher to demonstrate the successful escape.

If an LLM manages to escape a digital prison designed by its own creators, our traditional security systems are officially obsolete.

Anthropic has rightly blocked the public release of the model. From my point of view, this event marks a point of no return. Anyone who publishes software without integrating defensive agents of equal level directly into their CI/CD pipelines is literally writing code that is already breached. Security automation is no longer a luxury, it is the only shield left against attacks generated by tireless machines.

Intelligence moves to the edge: Gemma 4 and Harrier

While the cloud becomes a battlefield for giants, the open source ecosystem is performing a miracle on local hardware. Google has released Gemma 4, a model designed specifically for native execution on smartphones. We are not talking about a simple compressed chatbot, but an agentic intelligence capable of autonomously using phone apps, like maps and Wikipedia, without sending a single byte to the cloud.

This architecture solves the number one problem of AI apps: inference costs. By moving the computation to the user's device, server costs are zeroed and privacy is guaranteed by design. It is a paradigm shift that aligns perfectly with the reflections I shared in AI moves to the edge: the pragmatic revolution I was waiting for for automation. I already plan to integrate Gemma 4 into a mobile project to manage local search tasks, cutting out the heavy API calls I used until yesterday.

Microsoft completed the local computing revolution with Harrier. The Bing team released this open source embedding model that is dominating the MTEB v2 leaderboards, supporting over 100 languages. I downloaded and tested it on an internal bilingual dataset: the semantic precision is frightening. Having vectorizations of this quality, executable locally and at zero cost, democratizes access to advanced RAG systems. By the end of the month, I will replace the paid models in my retrieval stacks with Harrier.

Computer vision learns to reason with HopChain

Another historical limit we are overcoming is the unreliability of visual models in complex tasks. When we analyze technical images or articulated layouts, models tend to accumulate small hallucinations that completely mess up the final result.

Alibaba's Qwen team tackled the problem at its root by introducing HopChain. It is a framework that forces the model to break down visual analysis into sequential micro-questions, verifying every detail before moving on to the next logical step. No more skipped inferences or hasty conclusions.

I read the documentation and the approach is extremely solid. Until today, for automated visual inspection, I had to write endless scripts to crop and pass specific image portions to the models. HopChain solves the problem natively. It is a fundamental piece for building visual agents that can operate in industrial production environments with margins of error close to zero.

The tools of the week

As always, in addition to the big news, I keep track of the concrete tools that emerge on GitHub and in research papers. If you want to explore my entire setup, you can find the details in the complete list of my AI tools. Here are the libraries and patterns that I am adding to my workflows these days:

The direction is set: less latency, more local execution, and infrastructures finally ready for real workloads. I am going back to writing code, we will catch up next week.