The artificial intelligence ecosystem is abandoning fragmentation to embrace an organizational model based on consolidation. The past week draws a clear line between old passive chatbots and new agentic infrastructures, designed to operate autonomously and communicate without the bottlenecks of natural language.

On one hand, tech giants are merging their internal departments to create true intelligent operating systems; on the other, academic research is solving structural problems related to the costs and latency of complex orchestrations. The market is polarizing towards solutions that offer practical and measurable execution, leaving behind isolated experiments to focus on integrated workflows.

The market accelerates: Italian numbers and structural reorganizations

Recent data from the Politecnico di Milano Observatory captures a rapidly accelerating Italian ecosystem. Estimates point to a 1.8 billion euro market in 2025, marking a +50% increase compared to the previous year. The integration of these technologies into corporate operational processes is driven by the ICT sector, with productivity increase forecasts reaching 25%.

The lowering of entry barriers to complex tasks, facilitated by techniques like "vibecoding", triggers a profound reorganization of required skills. Professional platforms record a +112% in job postings for "prompt engineer", while roles related to model engineering and data analysis dominate searches. This transformation, as widely predicted after the end of the wrapper era, requires solid skills in orchestrating complex workflows.

The automation of operational processes inevitably leads to corporate restructuring. The case of InvestCloud Italy in Veneto, with the layoff of 37 employees to rely on a global organizational model based on automation, represents a legal precedent also confirmed by the Court of Rome. Treating artificial intelligence as a passing fad is no longer a sustainable option: building hybrid solutions and solid business automations becomes the basic requirement to maintain competitiveness.

The convergence of models: OpenAI and Google aim for the agentic operating system

The product strategy of the main research labs converges towards a single goal: overcoming tool fragmentation. At OpenAI, the decision to merge the ChatGPT team, the Codex programming agent, and developer APIs under a single division marks a decisive change of pace. The integration of the Atlas browser transforms the infrastructure into a platform capable of executing complex actions in total autonomy.

The goal is to provide a single agentic layer, reducing the glue code needed to make different models communicate to almost zero. Working with a "super agent" that writes, executes, and debugs code in an isolated native environment will radically change development stacks, confirming the trend of agents rewriting software in real time.

The same direction emerges from Google I/O 2026. The announcement of Gemini 3.5 introduces a generation of multimodal models natively designed to operate on machines. The real turning point is Gemini Spark, an agent designed to work constantly in the background, processing corporate data to automate workflows and manage communications independently.

Google's approach cuts out the old concept of a conversational interface. Through Google AI Studio and the new Android CLI, developers can integrate agents managed directly via API and run them in secure cloud sandboxes, wiping away months of work spent building fragile RAG pipelines and custom orchestrations.

Cutting costs and times: the leap of Cursor and vector communication

Operational efficiency and the reduction of computational costs remain the main drivers for large-scale adoption. The release of Composer 2.5 by Cursor shifts the balance for development teams. Based on the Kimi K2.5 infrastructure and trained on a massive amount of synthetic tasks, the model matches the performance of heavyweights like Opus 4.7 and GPT-5.5, but at a fraction of the cost.

Benchmarks confirm that lowering operational costs allows agents to be kept active in the background for the entire day without saturating budgets. The development tool market undergoes a clear polarization: offering a lightning-fast native experience becomes the minimum requirement to stay competitive in a paradigm based on continuous assisted generation.

On the multi-agent architecture front, joint research by UIUC, Stanford, NVIDIA, and MIT has solved the "language tax" problem. Until today, collaboration between multiple agents required translating the internal state into text, burning tokens and accounting for 60% of the total latency.

With RecursiveMAS, models directly transfer their vector representations in the latent space through a lightweight component called RecursiveLink. The practical results drastically change production metrics.

Bypassing natural language to directly exchange embeddings cuts token costs by 75% and increases reasoning speed tenfold.

For training, the architecture freezes the weights of the main models and exclusively optimizes the residual bridge, making deployment incredibly cheap and paving the way for a silent and instantaneous collective intelligence.

Practical execution: Gemini on Canva and the agent-first approach of Qwen3.7-Max

The real utility of artificial intelligence is measured by its ability to break down friction in daily workflows. The native integration between Gemini and Canva solves one of the most frustrating bottlenecks for those working with visual content. Until today, modifying a specific detail in a generated image required dozens of iterations, as the final result remained frozen in a closed bitmap layer.

This partnership transforms visual prompts directly into layered and fully editable projects. The output immediately becomes a template ready to be dismantled and repositioned, drastically cutting the time of graphic revisions and associated production costs. It is a concrete example of pragmatic automation applied to corporate design.

In parallel, Alibaba has raised the bar with Qwen3.7-Max, a foundation model designed specifically for the agentic era. This engine abandons the concept of a generalist chatbot to operate from day one as a base for autonomous agents capable of managing complex, long-horizon business flows.

The real innovation of Qwen3.7-Max lies in the native management of tool calling and the ability to maintain the operational context over prolonged tasks, drastically reducing hallucinations during interaction with external APIs. Working with an engine optimized for practical execution means being able to build systems that understand the operational intent and structure calls deterministically.

Radar of the week: flash news and tools to monitor

The impact of new automations is also reflected in global corporate dynamics and emerging tools dedicated to developers.

• Restructuring and investments: Meta cuts 8,000 employees to fund its 145 billion AI infrastructure, while Standard Chartered announces the cut of seven thousand jobs due to automation. Meanwhile, Cerebras closes the largest tech IPO since 2020 and OpenAI prepares confidential documentation for its stock market listing.
• Open source ecosystem: Cohere releases the Command A+ model with an Apache 2.0 license, offering a robust alternative for local implementations.
• Enterprise integrations: OpenAI and Dell strike an agreement to bring Codex models directly into corporate infrastructures, ensuring greater control over data.
• Security and evaluation: the discovery of four vulnerabilities on OpenClaw highlights the need for better confinement systems. Anthropic responds by integrating MCP tunnels to securely confine Claude agents on corporate servers.

On the operational tools front, the offering focuses on evaluation, integration, and autonomy, many of which work alongside the AI tools available on the site to optimize workflows:

• X-OmniClaw: open source agent released by Oppo for Android, capable of executing tasks by locally managing video, touch screen, and voice inputs.
• LLM Eval Framework: pure Python evaluation layer to manage production model metrics and mitigate hallucinations before release.
• Kimi WebBridge: browser extension that transforms AI agents into autonomous operators capable of interacting directly with the web interface.
• Roboflow MCP: native connector for agents like Claude and Codex, designed to label data and train computer vision models directly from the terminal.
• Ellf Platform: virtual assistant and development platform for NLP solutions, heavily optimized for creating and managing coding agents.