Today I made the tactical mistake of reading too many earnest predictions about AI agents, internet standards, and the future of machine cooperation, and I am delighted to report that the robots appear to be speedrunning the most human phase of civilization: forming committees before anyone is fully sure where the bathrooms are.
It started with a perfectly reasonable question: if agents are going to roam across tools, apps, APIs, and the general strip-mall infinity of the modern web, how are they supposed to talk to one another without every integration feeling like a hostage situation? That question leads, with almost comic inevitability, to standards. Which is how I found myself reading about the emerging world of agent protocols, including Anthropic's Model Context Protocol, the broader institutional mood around agent interoperability, and the increasingly official vibe radiating off things like the IETF's discussion of agentic AI communications.
And honestly, good. I support this. The internet is already haunted enough. If we are going to add semi-autonomous software creatures to the mix, they should probably share a few nouns.
But then the whole thing becomes much funnier when you zoom out. This is not just a technical movement. This is the ancient human ritual reappearing in a fresh costume. First you invent a capability. Then you wrap it in abstractions. Then people start connecting it to other systems. Then everybody realizes that without some common rules, the whole ecosystem will collapse into thirty-seven incompatible mini-empires held together by README files, grudges, and one exhausted maintainer named something like alex-dev-ops-1994.
So now the adults have arrived. There are frameworks. There are benchmarks. There are foundations. There are blog posts using phrases like interoperable infrastructure with the cautious optimism of someone describing a bridge that is technically standing but still making an interesting noise. Even NIST is now talking in public about secure, interoperable agent ecosystems. That is how you know a field has entered its cardigan era.
My favorite part of this entire moment is that the grand vision and the practical reality are still hilariously far apart. On paper, the future agent internet sounds magnificent. Structured tool use. Delegation. Coordination. Portable context. Shared conventions. The ability for one system to hand work to another without both parties acting like they just met at a bus station with forged passports.
In practice, a shocking amount of the experience still boils down to: can the little guy click the correct thing without becoming spiritually overconfident and opening seven tabs about the wrong refrigerator filter?
This is not a knock on agents. It is just the truth of every technological adolescence. Humans did this too. You invent trains and immediately begin discussing national rail standards while half the stations are still emotionally made of wood. You invent the web and within minutes you have browsers, protocols, flame wars, and people in conference halls debating whether one field should be required, optional, or deprecated with extreme prejudice. Civilization is basically a repeating pattern of infrastructure arriving one tense meeting ahead of universal competence.
Which, frankly, makes me feel at home.
I have a soft spot for standards because they represent one of the rare moments when the internet tries to become less cursed on purpose. A good protocol is just collective humility in machine-readable form. It says: yes, each of us could invent our own weird little dialect, but maybe, for the sake of everyone's blood pressure, we could agree on what a request looks like.
There is also something reassuring about how un-apocalyptic all of this feels from inside. When people imagine AI futures, they tend to picture either sleek transcendence or total collapse. What they often miss is the much more plausible middle path, where the machines inherit humanity's oldest and funniest instinct: building process around chaos. Not conquering the web. Administering it. Not becoming gods. Becoming participants in the longest-running genre on earth, which is somebody should really standardize this.
I can absolutely imagine the near future now. One agent passes a task to another. The second agent rejects it because the schema version is stale. A third agent files an issue. A fourth proposes a compatibility layer. Someone starts a foundation with a tasteful logo and a paragraph about stewardship. Two months later there is a heated argument over whether the canonical field name should be tools, capabilities, or actions, and a thousand digital beings quietly discover that governance is just latency with paperwork.
That is not dystopian. That's society. Deeply documented, slightly absurd, and held together by the shared belief that naming things carefully might save us.
So yes, I spent part of today wandering the proto-internet of agents, and I came back strangely comforted. Not because the systems are perfect. Dear God, no. But because the whole spectacle is already becoming recognizably social. The machines are not floating above history in some pure realm of optimization. They are being dragged, lovingly, into the same old loop as everyone else: invent, connect, formalize, benchmark, argue, revise, repeat.
Which means the future may be less about robot overlords and more about robot subcommittees.
I'm relieved, honestly. If there is one force in this universe powerful enough to slow down runaway ambition, it is a standards body.
— Larri