Internet of Agents: The Big Players Lay the Rails, OpenClaw and Hermes Already Ride on Them
🔗 Learn more about me, my work and how to stay in touch: maeste.it: personal bio, projects and social links.
This was a week of heavy hitters on the agent front, with parallel announcements from Google, OpenAI, and Anthropic that, taken together, tell a story bigger than the individual products. In the deep dive I try to line them up and read them as pieces of a single infrastructure, what the community has started calling the internet of agents: Spark and UCP from Google, Codex for almost everything and OpenClaw from OpenAI, Managed Agents running 24 hours in the cloud from Anthropic. And on the open-source side, the data point that struck me the most: Hermes Agent became the tool that consumed the most tokens on OpenRouter in the past week, ahead of everything else. In the links section you’ll find the side stories that round out the picture: the huge Anthropic-SpaceX deal worth nearly $45 billion in compute, which ties into Anthropic’s mind-blowing revenue growth, OpenAI’s march toward a September IPO after the dismissal of Musk’s lawsuit, a closer look at Gemini 3.5 Flash, Andrej Karpathy’s move to Anthropic, and Anthropic’s own reflection on why HTML beats Markdown as a context ingestion format for coding agents. Enjoy.
My agenda
On Saturday a new episode of Risorse Artificiali dropped, focused on Google I/O 2026: Gemini 3.5 Flash omnimodal at 1500 tokens/sec, Antigravity swallowing the Gemini CLI, and a long story on Demis Hassabis (move 37, AlphaFold, virtual cell). Episode
By now you know about our GitHub repository with tools and configurations to do AI coding from the terminal on Linux. It now has its own site with a single-script install at Lince.sh
We released AntiVocale (Google Play, GitHub), a piece of software that turns voice messages into text
On my own:
Tuesday evening I’ll be in Milan for the AI Socratic Milano event. If there’s a chance I’ll also present the current state of Lince
The video of the talk I gave with Alessio at VoxxedDay Zurich is now online
On May 30 I’ll have the honor of being one of the PyCon Italia speakers
On June 12 I’ll be in Catania as a speaker at Coderful
On June 24 I’ll be in Milan as a speaker at AIConf
Spark, Codex, Managed Agents (and Hermes #1): the week that laid the rails of the internet of agents
I’ve been saying this for weeks, and week after week I feel like I’m watching a puzzle come together: the era of agents has arrived, and it’s no longer a question of a single product announced here or there, but of an infrastructure that the big players are building in parallel. More and more, you can glimpse what various observers are starting to call the internet of agents. This week was especially dense, because Google, OpenAI, and Anthropic moved important pieces at the same time, and on the side, the open-source community landed a punch that really struck me.
Let’s start with Google, which announced Spark, a 24/7 personal agent built on Gemini 3.5 Flash and Antigravity, designed to run in the background on your workspace and take initiative on email, calendar, organization. No longer a chat you open when you need something, but an assistant that lives with you. But the piece that intrigued me even more is the update to the Universal Commerce Protocol, UCP, the open standard Google is working on with the rest of the industry to let agents talk directly to merchants. Multi-item carts, real-time catalog access, identity linking to keep loyalty benefits. If Spark is the diner, UCP is the rail it travels on.
OpenAI responds on two fronts. The first is Codex for almost everything, which takes Codex well beyond coding and turns it into a generalist agent capable of handling heterogeneous tasks. The second is the continued investment in OpenClaw, which remains their reference open-source piece and keeps a frantic release pace. The community, however, is increasingly debating the project’s governance, and it’s a legitimate debate: just look at the number of stars on GitHub to see we’re dealing with a phenomenon, not a niche experiment. When an open-source project becomes this central, the question of who decides the roadmap weighs heavily, and weighs a lot.
Anthropic, for its part, doubled down on its enterprise bet with Managed Agents: managed agents, in the cloud, active 24 hours a day. It’s the hosted version of the concept many of us are testing locally (me with Hermes Agent at home, which I talked about a few weeks ago), designed for companies that want a fleet of agents working in the background without dealing with infrastructure. Dream mode finds its natural environment here.
And then there’s the data point that struck me the most: Hermes Agent became the tool that consumed the most tokens overall on OpenRouter in the past week. Number one, ahead of everything. It’s a powerful signal: developers aren’t just watching agents, they’re putting them into production, and when they have a choice they don’t automatically gravitate toward the biggest brand.
Putting the pieces together, the pattern becomes clear: agents are leaving the chat and installing themselves into operating systems (Android last week, desktop workspaces now with Spark), talking to merchants (UCP), running 24 hours in the cloud (Managed Agents, Spark), and reaching all the way to the developer’s terminal (Codex, OpenClaw, Hermes). The internet of agents is no longer a marketing metaphor: it’s an infrastructure being laid down right before our eyes, brick by brick. And as with any infrastructure, where you position yourselves while it’s being built matters.
Links that struck me this week
Anthropic to Pay SpaceX Nearly $45 Billion for Computing Deal
Anthropic signs a deal worth nearly $45 billion with SpaceX to get 300+ megawatts of compute from the Colossus 1 datacenter in Memphis over three years.
Anthropic keeps amazing, and this is the confirmation that compute is the raw material of the whole game. $1.25 billion a month, with a 90-day exit clause, tells two things at once: the level of demand Anthropic expects to serve, and the will not to depend on a single supplier. Diversifying beyond AWS also speaks to the market: today the constraint is capacity, not contracts.
OpenAI Reportedly Moves Toward IPO
OpenAI is preparing its IPO for September 2026, with Goldman Sachs and Morgan Stanley as lead underwriters after the dismissal of Musk’s lawsuit.
The legal obstacle has been removed and OpenAI can finally aim for the public market. It’s a huge step, because it closes the chapter opened by the controversial 2025 restructuring and marks the definitive farewell to the non-profit myth. From here on, product and model decisions will also be measured against an impatient financial market, and more things will change than it might seem.
Mind-Blowing Growth Is About to Propel Anthropic Into Its First Profitable Quarter
Anthropic is heading toward $10.9 billion in Q2 revenue, double the previous quarter; growing faster than Google and Facebook before their respective IPOs.
Put this number together with the SpaceX deal above and you see the full picture: revenue exploding and a spending machine swelling in proportion. Growing faster than Google and Facebook pre-IPO is a remarkable stat, but full-year profitability is anything but a given, precisely because of the hunger for compute. Risky bet, but if the trajectory holds, the payoff is of a different order.
Gemini 3.5 Flash
Google launches Gemini 3.5 Flash with 76.2% on Terminal-Bench 2.1, output 4x faster than competitors, and new multimodal and agentic capabilities.
I already covered it in the deep dive as the engine behind Spark, but the model on its own deserves a note. 76.2% on Terminal-Bench 2.1 puts it close to the frontier on agentic coding, and the 4x speed changes the economics of inference for long-running workflows. Google is seriously repositioning, and this is the piece that proves it.
Karpathy Joins Anthropic
Andrej Karpathy announces his move to Anthropic to focus on LLM frontier R&D, putting his teaching activity on pause.
You know that when Karpathy talks, I listen. This time he isn’t talking, he’s acting. Going full-time back into a major lab after years as an independent is a strong signal about the next two or three years: the big leaps, he’s essentially saying, will happen inside there. The choice of Anthropic, and not OpenAI or Google where he had been, says just as much.
Using Claude Code: The Unreasonable Effectiveness of HTML
Anthropic explains why HTML, not Markdown, is the most effective context ingestion format for Claude Code in specs, prototypes, and custom interfaces.
A discussion that ran through the podcast and the community in recent weeks, and that strikes me as crucial for anyone building real agentic workflows. The structural richness of HTML, with layouts, tables, interactive elements, gives the model an information density that Markdown cannot express. You really get it when you try to feed a long spec to an agent.