AI Weekly Trends Highly Opinionated Signals from the Week [CY26W11]

🔗 Learn more about me, my work, and how to connect: maeste.it – personal bio, projects, and social links.

The usual intense week with Meta, which further delays the release of its new LLM models, while acquiring Moltbook, effectively certifying that we are heading towards an agent-based economy. Professor Ethan Mollick, whom I have cited many times in this newsletter and whom I greatly respect, also writes about how the AI world is already shifting from what he calls co-intelligence to a model where humans use their own intelligence to orchestrate multiple agents that do the heavy lifting. It’s a significant and in some ways epochal shift, especially coming from someone who has always strongly theorized the concept of co-intelligence.

But there are many other news stories and many other new tools we discuss in this newsletter, in addition to the fact that I’ve reintroduced a section for this issue where I discuss the main research papers related to the trends I highlighted earlier. It’s something else I’ve been thinking about in recent weeks, namely persistent memory, skills, harness development and how agents are increasingly viewed as entities for reinforcement learning, beyond the single model they use, as dispositive entities.

Before leaving 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 willing to do so):

• Podcast with Alessio and Paolo:

  • On March 12th we were at the JUG in Milan to record our first live episode.

  • We are working on more interviews and episodes with very interesting guests

  • We created a GitHub repository with tools and configurations for AI coding from the terminal on Linux. Obviously open source, so check it out and contribute: LINCE - Linux Intelligent Native Coding Environment

• Solo:

  • On March 19th I’ll be at the AI aperitivo, not sure yet if I’ll be doing a Lince demo... but we can talk about it regardless :)

  • 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

AI Models News and Research

Takeaways for AI Engineers

• Takeaway 1: Competitive value is shifting from the model to the integrated product.

• Takeaway 2: Meta confirms its post-Llama 4 struggles: the Avocado delay signals a growing gap.

• Takeaway 3: Code generation is evolving towards complete product generation.

• Action Items:

• Assess how much your workflows depend on isolated models vs. integrated products.

• Explore the new Gemini integrations in Google Workspace and Maps to understand the level of maturity achieved.

What’s happening this week?

In this week’s model news (or rather, more product news than model news) I’m focusing on slightly different aspects than what I normally pay attention to. I usually follow model releases closely, but this week the only model news is rather negative: Meta further delays the launch of the new Avocado models and generally what are essentially the Llama 5 models. This release is delayed at least until May, and what’s leaking are performance and accuracy issues. Meta really isn’t making a good impression on the market since Llama 4.

The other news instead focuses on what is a fairly strong trend: the shift from isolated systems to much more integrated systems, where models collaborate directly with products. In this regard, Google’s updates are noteworthy, both for Maps and for Workspace updates, where Gemini models are deeply integrated with Google’s products.

Last but not least, the mention of Replit Agent 4, which I cite here rather than in the AI Assisted Coding section because I believe the shift from pure code generation to collaborative generation of an entire product suite is significant.

Links of the Week

• How We’re Rethinking Maps with Gemini — Google’s Ask Maps uses Gemini for personalized real-time answers and destination recommendations.

• Claude Now Creates Charts, Diagrams and Interactive Visualizations — Imagine with Claude generates and edits charts, diagrams and interactive visualizations directly in conversation.

• Meta Delays New AI Model Release Over Performance Concerns — Meta’s Avocado model doesn’t compete with leaders; release postponed at least to May over performance issues.

• Gemini Workspace Updates — New Gemini features integrated in Docs, Sheets, Slides and Drive for enhanced productivity and collaboration.

• Replit Agent 4 — Infinite design canvas and parallel AI agents to build backend, frontend and slide decks in a single integrated environment.

Agentic AI

Takeaways for AI Engineers

• Takeaway 1: The A2A protocol v1.0 marks the shift from isolated agents to interoperable, production-ready multi-agent ecosystems.

• Takeaway 2: Karpathy’s AutoResearch shows agents iteratively improving models: a concrete step toward autonomous self-improvement.

• Takeaway 3: Three architectural patterns are consolidating for agents: persistent memory, programmable skills and harnesses as autonomy infrastructure.

• Action Items:

• Study the three patterns (memory, skills, harness) and assess which are already present in your agentic architecture.

• Explore the A2A protocol v1.0 and its Agent Cards to understand how to enable cross-platform communication between your agents.

What’s happening this week?

In this section there’s truly an embarrassment of riches when it comes to choosing which news to focus on, but I can’t help but start with the announcement of the A2A protocol in version 1.0, since I actively participated in this work with my team, also maintaining the Java SDK and TCK for the entire protocol.

But leaving personal matters aside, an honorable mention goes to Karpathy’s AutoResearch, which has been making waves in the community for about two weeks. It involves using agents in a guided research loop to improve a model’s training. I’ve said it many times, both here and on the podcast, that seeing models capable of improving themselves brings AGI to mind very closely. For me, that’s one of the turning points. We may not be there yet, but seeing iterative improvements, managed entirely by an agent, is certainly impressive.

Then there are three trends I’ve been highlighting for some time. Persistent memory for agents, skills as a way to extend agents and program their behaviors, and finally a trend I’ve been emphasizing for a few weeks now: harnesses. By harness we mean all those artifacts (tools, agents, memory or anything else that can be used by the LLM) to behave as autonomously and decisionally as possible, to become an agent. I’m including an article for each of these trends, which I consider significant cultural knowledge for any AI engineer.

Finally, Perplexity demonstrates that the idea of letting agents use a computer directly is realistic and not just a toy made by the community like OpenClaw. It’s a bit like when we talk about humanoid robots that have that form factor to be able to use all the tools we designed for the human form factor. In the same way, agents capable of using existing software, even if designed for human use, can have the competitive advantage of reusing a huge base of already available tools.

Links of the Week

• Agent Skills: Progressive Disclosure as a System Design Pattern — Three-level pattern (discovery, activation, execution) for managing agent context efficiently.

• A2A Protocol v1.0: Standardized Agent Communication — Open protocol for discovery, communication and coordination between AI agents across different platforms and organizations.

• The Anatomy of an Agent Harness — Models contain intelligence, the harness makes it useful: core components for transforming LLMs into agents.

• Google Always On Memory Agent — Open source system for persistent agent memory, without vector database, under MIT license.

• Karpathy’s AutoResearch — AI-driven research loops for iteratively improving model training on a single GPU.

• Perplexity’s Personal Computer — AI agents managing tasks by delegating to other AIs on a Mac Mini, like an automated project manager.

AI Assisted Coding

Takeaways for AI Engineers

• Takeaway 1: 2026 is shaping up as the year LLMs enter code review: Claude Code Review is the first concrete signal.

• Takeaway 2: The extension ecosystem for Claude Code is maturing rapidly, with projects like Everything Claude Code and SuperClaude enriching the agentic experience.

• Takeaway 3: Chrome DevTools MCP opens browser functionality to any agent, not just coding ones, with official Google support.

• Action Items:

• Try Claude Code Review on your PRs to evaluate the quality of automated reviews compared to the manual process.

• Configure Chrome DevTools MCP in your development environment to explore agentic debugging possibilities.

What’s happening this week?

There are memes in the community about how you go to bed, wake up and every morning find a new Claude feature, and they pretty much reflect reality. Just this week there are at least three major announcements in the Anthropic world. The first and most important is that they released in preview for Team and Enterprise customers a new Claude Code feature called Review for managing pull request reviews. Just last week I was talking about how LLM usage is shifting from pure code generation to other phases as well. 2026 could be the year we see LLMs become a significant part of the code review toolchain. This seems like the first signal.

The second Claude link I’m highlighting is called Everything Claude Code and it’s interesting because it’s one of the winners of the Anthropic hackathon. It’s a series of agents, commands and skills designed to improve the Claude Code experience, something very similar to SuperClaude which I’ve already discussed many times. It’s on my to-do list this week to try it extensively to see if the experience is genuinely better than SuperClaude and other already available projects.

Meanwhile at Google they’re announcing Chrome DevTools as MCP, and it’s an important announcement because beyond the development aspect it allows opening browser functionality to any agent, coding or not. Something we’ve certainly already seen happen with OpenClaw, but this time with full Google support.

Links of the Week

• Claude Code Review — Automated code review system with multi-agent team for in-depth pull request analysis, available for Team and Enterprise.

• Everything Claude Code — 16 specialized agents, 65+ skills and 40+ slash commands to optimize Claude Code workflows.

• Claude Interactive Mode Documentation, /btw — Side questions during active work without interrupting tasks or adding to conversation history.

• How Coding Agents Are Reshaping Engineering, Product and Design — The bottleneck shifts from writing to reviewing code; generalists gain the most advantage.

• Chrome DevTools MCP — Direct connection to active browser sessions for agentic debugging, no extensions or headless browser needed.

Business and Society

Takeaways for AI Engineers

• Takeaway 1: Meta’s acquisition of Moltbook confirms big tech’s interest in an agent-based economy.

• Takeaway 2: The Anthropic Institute signals that legal, economic and AI governance challenges are becoming strategic priorities on par with technical development.

• Takeaway 3: Ethan Mollick redefines the relationship with AI: from co-intelligence (AI helping humans) to AI management (humans orchestrating autonomous agents).

• Action Items:

• Read Ethan Mollick’s full article “The Shape of the Thing” to deepen your understanding of the shift from co-intelligence to AI management.

• Watch OpenAI’s open source acquisition strategy as an indicator of which tools will become platform standards.

What’s happening this week?

In this section we start again from Meta, which acquired, or rather hired, the person who created Moltbook, since it’s an acquisition of a company made by a single person. For those who don’t remember, Moltbook is a social network for agents that seems to be exactly in Meta’s core business, but also confirms the interest from major companies in an agent-based economy.

Meanwhile OpenAI continues with its acquisitions of open source platforms. Just as happened with OpenClaw, at least in acquiring open source platforms and keeping them open it’s living up to its name that starts with Open, something it has never done, or almost never, with its models.

And while Nvidia is also investing in Mira Murati’s startup with a multi-year partnership, I prefer to focus on two articles. One from Anthropic, presenting the Anthropic Institute, an institute founded to focus on the legal, economic and global governance aspects of AI. And one from Professor Ethan Mollick, who examines AI’s transition from what he has always called co-intelligence, meaning intelligence augmented by humans through the use of AI, towards instead an intelligence that must be used by humans for managing AI and the many agents that can be used in parallel to accomplish complete tasks. I recommend reading this article carefully because Professor Ethan Mollick certainly expresses this concept better than I can, and I certainly don’t want to try to summarize it when I think a complete and thorough reading is essential.

Links of the Week

• The Shape of the Thing — Ethan Mollick examines the shift from co-intelligence to AI management and autonomous agents.

• Introducing the Anthropic Institute — Institute dedicated to legal, economic and global AI governance aspects, led by Jack Clark.

• Nvidia Invests in Mira Murati’s Thinking Machines Lab — Multi-year partnership with at least one gigawatt of chips for frontier model training and serving.

• Promptfoo Joins OpenAI — Open source AI safety and evaluation platform acquired by OpenAI, will remain open source.

• Meta Acquired Moltbook — Meta acquires the AI agent social network built on the OpenClaw framework.

Research Papers

Takeaways for AI Engineers

• Takeaway 1: The papers confirm the agent harness trend: skills, tools and memory are the infrastructure transforming AI from informative to dispositive.

• Takeaway 2: Reinforcement learning applied to agents (not just LLMs) paves the way for enterprise agents trained for specific use cases.

• Takeaway 3: SkillNet proposes an open model for reusable and composable skills, a useful reference for anyone designing agentic architectures.

• Action Items:

• Read SkillNet for inspiration on how to structure and connect skills in your agents.

• Explore the distinction between RL on LLMs and RL on agents introduced by KARL to understand the implications for your system design.

What’s happening this week?

A research papers section that was missing from this newsletter for a while. I’m reintroducing it to bring you four very significant papers I’ve read in recent weeks that confirm the trends highlighted in the previous sections: the need for long-horizon memory (Memex(RL)), with reinforcement learning that allows agents to use it better, but also the entire trend developing around agent harnesses (AutoHarness), namely skills, tools and everything related to the iterative refinement of the instruments agents can use to transition from an informative artificial intelligence to a dispositive one.

In the same direction goes SkillNet, a paper that defines an open infrastructure for reusable AI skills, with multidimensional evaluation and connections between them. It’s worth reading even just for ideas on how to better write and connect your own skills.

The KARL paper discusses knowledge agents via reinforcement learning. It’s interesting for demonstrating that through reinforcement learning, agents, particularly research ones but not only, can be trained for specific enterprise cases as distinct entities for clients. Here we’re talking about applying reinforcement learning to agents, meaning the reasoning and action part, and not just the LLM part. It’s an important distinction.

Links of the Week

• AutoHarness: Automatic Code Harness Synthesis for LLM Agents — Automatic generation of protective structures for agents through iterative refinement with environmental feedback.

• SkillNet: Creation, Evaluation and Connection of AI Skills — Open infrastructure for reusable AI skills with multidimensional evaluation and 200,000+ skills in the repository.

• KARL: Knowledge Agents via Reinforcement Learning — Enterprise research agents trained with RL, with Pareto-optimal performance compared to Claude 4.6 and GPT 5.2.

• Memex(RL): Scaling Long-Horizon LLM Agents via Indexed Memory — Indexed memory with RL for long-horizon agents, overcoming the limits of finite context windows.