AI Weekly Trends Highly Opinionated Signals from the Week [CY26W12]
🔗 Learn more about me, my work, and how to connect: maeste.it – personal bio, projects, and social links.
If I had to find a recurring theme in this newsletter, and perhaps in those of the past few weeks as well, I believe it’s undoubtedly agentic coding and more broadly the development of agentic systems. But for now, not distributed across the network as one might have thought at the beginning of this development, rather very much centered on the user’s machine. This will probably change over time and, as we’ve seen many times, cycles and recycles within computing: it has often happened in the past that a centralized solution ended up on individual users’ machines only to return to distributed solutions.
Think about it: we went from mainframes to personal computers and then back to the cloud, but this is just one of the many examples I could give of this trajectory, and it’s what is somewhat happening in the world of AI as well. We went from chatbots, centralized and completely controlled by the vendor, to coding agents or personal agents (think not only of Claude Code, but also OpenClaw or Claude Work), and then perhaps in the future we’ll see agents distributed in the cloud. The latter hasn’t happened yet, but all the conditions are there for it to happen. So keep your eyes open, keep reading the newsletter and always maintain your own critical thinking about what’s happening. Get your hands dirty trying some of the things I suggest and do everything you can to jump on this fast-moving train.
Before I leave you to the news and my analysis of what happened this week, let me share what has happened, is about to happen, or will happen in my public agenda, for those who want to follow my talks or meet me in person (I love exchanging opinions with anyone who’s willing):
Podcast with Alessio and Paolo:
On Wednesday, my interview with Massimo Re Ferré, PM at AWS Kiro, was released.
We’re working on more interviews and episodes with very interesting guests.
You already know about our GitHub repository with tools and configurations for AI coding from the Linux terminal. This week we released a complete dashboard... almost an IDE for agents, but all from the terminal: LINCE - Linux Intelligent Native Coding Environment
On my own:
On March 24th I’ll be at Voxxed Day in Zurich. Alessio and I are presenting a talk on AI assisted coding
On March 25th I’ll be speaking at this meetup in Milan on Vibe Coding and Agentic Engineering
On May 30th I’ll have the honor of being one of the PyCon Italia speakers
On June 12th I’ll be in Catania as a speaker at Coderful
AI Models News and Research
Takeaways for AI Engineers
Takeaway 1: Anthropic’s 1M token context qualitatively changes working with coding agents, eliminating the context management bottleneck
Takeaway 2: Chinese vendors are converging on models optimized for agentic scenarios (MiniMax M2.7, GLM-5-Turbo): agentic AI is the competitive battlefield of the moment
Takeaway 3: If DeepMind is building metrics for AGI, it’s because the perceived distance is concretely shrinking
Action Items:
Try Claude Code with 1M context on a complex multi-file task
Read the DeepMind paper on the cognitive framework for AGI to calibrate expectations on where we really are
What’s happening this week?
Let’s start with a relevant piece of news from Anthropic, which has extended the context length of both the Opus and Sonnet models to one million tokens. Having one million tokens significantly changes the experience of using Claude Code, especially because it allows you to perform very long and complex operations without the need to clear the context or keep track of these operations in alternative ways. It’s definitely a feature worth trying.
Chinese model vendors are certainly not standing still either. MiniMax launches its M2.7 model, which has excellent benchmark results especially for agentic functionality. The same goes for GLM-5-Turbo, also optimized for agentic scenarios. Both models are recommended for use with OpenClaw, which is, or at least seems to be, the real reference point right now for Chinese models. In practical use for coding, I said in previous weeks that the experience with Claude is still superior. But I must admit that I also use one of these models.
Very interesting instead are the news coming from research, starting with World Models that are advancing AI further by simulating the complexity of the real world. But also an interesting piece of research from DeepSeek on using an index within attention, better capturing the semantic meaning of the sentence being processed.
Last but not least, in the research space, a cognitive framework released by Google DeepMind to measure progress toward AGI. Beyond the paper, which is very interesting to read and I recommend, I’d like to emphasize that if DeepMind, with the visibility it has on model evolution, is designing a framework to measure progress toward AGI, it’s because we’re getting ever closer to that point.
Links of the week
1M context now available for Opus 4.6 and Sonnet 4.6 — 1M token context window at standard pricing for Claude Opus 4.6 and Sonnet 4.6, including Claude Code
World Models: Computing the Uncomputable — World Models simulate real-world complexity to enable prediction and planning through neural networks
Measuring progress toward AGI: a cognitive framework — Google DeepMind proposes a cognitive taxonomy with 10 key abilities to measure progress toward AGI
MiniMax launches M2.7 model — M2.7 model available via API with autonomous debugging capabilities and research agent harnesses
GLM-5-Turbo — Z.ai’s foundation model optimized for agentic scenarios with 200K context and MCP support
Faster Sparse Attention with IndexCache — Patch for SGLang and vLLM that eliminates up to 75% of indexer computations in DeepSeek Sparse Attention
Agentic AI
Takeaways for AI Engineers
Takeaway 1: OpenAI is building a complete ecosystem for agentic coding: model (GPT 5.4), security monitoring, and subagents in a single coordinated strategy
Takeaway 2: Subagents are now a consolidated cross-platform pattern (Claude Code, Codex): those not using them are leaving performance on the table
Takeaway 3: Context engineering is the key discipline for making agentic systems work well, and SwirlAI’s article maps the five fundamental patterns
Action Items:
Read the “State of Context Engineering in 2026” article as a practical guide for optimizing context in your agents
Try subagents in Codex or Claude Code to experience the agentic delegation pattern on a real task
What’s happening this week?
OpenAI demonstrates that it has developed great interest in coding and positions itself as a serious alternative, according to some even better, to Claude Code. This week I invite you to focus on the three articles I’m reporting from OpenAI, because all three are significant in this strategy.
GPT 5.4 has been a major step forward especially in its agentic usability within Codex and is the first model and agent, together with Codex, from OpenAI that truly seems capable of handling a wide variety of tasks, both code and non-code. Furthermore, OpenAI has released a monitoring system for autonomous coding agents, designed to detect misalignment risks and study their behavior, which further underscores their interest in this market. The same goes for the fact that Codex now supports subagents, a pattern widely used in the Anthropic world and beyond, which allows launching autonomous secondary agents managed by the main agent while saving context and optimizing work.
Moving away from the OpenAI world, I’d like to highlight OpenShell in this sector, a secure runtime made by NVIDIA for running autonomous agents. Essentially it’s a solution based on Kubernetes with declarative policies written in YAML. Truly an advanced solution that goes well beyond, probably, just writing code.
Lastly, I’d like to point out a very interesting article about context, called “State of Context Engineering”, which contains truly four or five very interesting insights. In my opinion it’s a must-read right now to deeply understand how to effectively manage context in your agentic systems. It ranges from skills and their progressive disclosure to context compression, intelligent routing, but also topics related to RAG or external tools like tools or MCP servers.
Links of the week
GPT 5.4 is a big step for Codex — First OpenAI agent with true agentic usability, precise instructions and ability to handle diverse tasks
Monitoring Autonomous Coding Agents — OpenAI monitoring system to detect misalignment risks in internal coding agents
Subagents and custom agents in Codex — Codex releases subagents in GA with support for custom agents defined via TOML files
OpenShell — Secure NVIDIA runtime for autonomous agents with sandbox and declarative YAML policies on Kubernetes
State of Context Engineering in 2026 — Five key patterns for managing context: progressive disclosure, compression, routing, RAG and MCP
AI Assisted Coding
Takeaways for AI Engineers
Takeaway 1: OpenAI acquires Astral (uv, Ruff, ty) and promises to keep them open source: the word “open” in their name might finally make sense at least for tools
Takeaway 2: Skills are the highest-impact extension point for coding agents: those not using them are underutilizing their agent
Takeaway 3: The coding agent market is fragmenting with serious alternatives (Cursor Composer 2, Codex) challenging Claude Code on price and performance
Action Items:
Read the Anthropic article on skills and start adding them to your coding agents
Try LINCE Dashboard to manage multiple agents in parallel from the terminal
What’s happening this week?
The acquisition of Astral by OpenAI speaks volumes about how much OpenAI has developed interest in the coding market. For those who don’t know, Astral is the company behind three of the main open source development tools for Python. I’m referring to uv, Ruff, and ty, which are now joining the Codex ecosystem. Just as they did with OpenClaw, OpenAI promises to keep the projects completely open source, which if confirmed would give meaning to the word “open” in their name, given that with models they have a completely different policy.
Cursor instead, which seemed to have been somewhat forgotten, releases its first truly frontier coding model. Composer 2, at a fairly low price, has substantial performance improvements that according to the vendor would even surpass those of Claude.
From Google comes the announcement of a new design tool called Stitch, which we had already seen peeking out in Google Labs, and which will transform the way of working and collaborating with AI in 3D and 2D environments, with the ability to generate functional React applications from designs alone, meaning not from code but from the application’s design itself.
One of Anthropic’s internal developers wrote a beautiful article on how skills are used internally at Anthropic and how they were developed. I strongly recommend reading this article if you’re adding skills to your toolkit for your coding agent. And if you’re not doing that, you should. So read that article, learn how to do it and start evaluating the addition of skills among the things you use to improve the coding experience with coding agents.
Finally, a mention from the personal front. Together with the other folks at Risorse Artificiali, this week we made available on GitHub a tool for using multiple agents efficiently and effectively within the Linux terminal. LINCE Dashboard has support for session persistence and also voice input capability. In the meantime, Anthropic has made voice input available on Linux as well, but trust me, ours works much better. LINCE Dashboard configures itself as a unique environment where you can use multiple agents even on different directories and therefore different projects in parallel, without ever missing an input request from one of them while you work in parallel on something else. Additionally, each agent is sandboxed to ensure security on your system, which I’ll never tire of repeating is one of the fundamental things when using a coding agent or any other kind of agent.
Links of the week
Lessons from Building Claude Code: How We Use Skills — How Anthropic uses skills internally: functional folders, progressive disclosure and high-impact “Gotchas” sections
Cursor Composer 2 — Frontier coding model at competitive pricing with substantial improvements on CursorBench and SWE-bench
OpenAI acquires Astral — uv, Ruff and ty join the Codex ecosystem; OpenAI promises to keep them open source
Early look at upcoming design tool from Google — Stitch: 3D workspace with AI that generates functional React applications from designs
LINCE Dashboard — Zellij plugin for managing multiple Claude Code instances with session persistence and voice input
Business and Society
Takeaways for AI Engineers
Takeaway 1: OpenAI is targeting an IPO by transforming ChatGPT into an enterprise productivity tool: the shift from consumer novelty to sustainable revenue is the real signal
Takeaway 2: China is democratizing AI with mass adoption programs that have no equivalent in the West
Takeaway 3: Open source in the AI era needs new mentorship frameworks to manage the noise from automatically generated contributions
Action Items:
Form your own opinion by reading the four links and share it in the comments
Evaluate how the “3 Cs” framework can apply to your open source projects that receive AI contributions
What’s happening this week?
In this section I invite you to read the four links I’m proposing without giving you too much of my own reading, because I believe it’s important that you have your own opinion and your own critical thinking on the links I propose. I’m happy to discuss any opinions in the comments, but I believe that the four links, while not being front-page news, are very significant for understanding which direction the world of AI is heading, with the United States and China leading the trends that shape the market.
Links of the week
OpenAI preps for IPO by end of year — OpenAI targets stock market listing by 2026, ChatGPT must become an enterprise productivity tool
How China is getting everyone on OpenClaw — Baidu and Tencent promote OpenClaw with mass adoption campaigns across all demographics
Rethinking open source mentorship in the AI era — The “3 Cs” framework for strategic mentorship against the noise of AI-generated contributions
Google’s Personal Intelligence expanding to all US users — Google’s assistant accesses Gmail and Photos for personalized responses, now available to everyone