For most of the short history of generative AI, we have imagined a fairly simple relationship.
A human asks a machine a question. The machine answers.
Even when the machine became an “agent”, the basic picture remained much the same. Give it a computer, some tools and a goal, and let it do things on our behalf.
That picture is already becoming obsolete.
The next important development in artificial intelligence may not be a much smarter model. It may be something less spectacular on the surface: AIs learning how to find, communicate with, employ and collaborate with other AIs.
Google’s Agent2Agent protocol, or A2A, is one attempt to create a common language for doing exactly this. It has now moved into the Agentic AI Foundation, placing it alongside a broader attempt to create open infrastructure for an agentic internet.
At first glance this sounds like another technical standard.
It isn’t.
Combined with increasingly capable open models, persistent memory, tools, shared workspaces and protocols such as MCP, it opens an extraordinary range of possibilities.
We may be beginning to build an internet whose participants don’t merely exchange information.
They think and act.
From AI assistants to AI societies
A2A allows independent AI systems to discover one another’s capabilities, exchange messages and data, undertake stateful tasks and return finished artifacts. Crucially, they don’t have to expose their internal models, memories or tools to one another.
That means an AI does not have to know how another AI works.
It merely needs to know what it can do.
Imagine asking your personal agent to investigate whether a particular solar system would work for your house.
Today we imagine one powerful model searching the web, doing the calculations and producing an answer.
An agent network could behave very differently.
Your agent might discover an electricity-pricing agent, a weather agent, a solar-engineering agent, a local planning agent and an economic forecasting agent.
Each investigates part of the problem.
Another agent checks their conclusions.
Your agent combines the resulting work and brings it back to you.
The interesting part is that nobody necessarily designed that team.
It formed because a problem existed.
Once the problem is finished, the team can disappear.
We have moved from using an intelligence to assembling intelligence.
Intelligence becomes heterogeneous
This also undermines one of the great assumptions of the current AI race: that the destination is an increasingly enormous general-purpose model capable of doing everything.
Perhaps it isn’t.
If agents can easily find and use other agents, specialisation becomes enormously valuable.
One node in the network might contain a frontier language model.
Another could run a tiny local model.
Another might contain a mathematical solver.
Another a database.
Another a physics simulation.
Another an industrial robot.
Another might simply contain a carefully written collection of rules.
And another could be a human being.
A2A is deliberately designed so that independent and potentially opaque systems can cooperate despite having different internal architectures.
The network therefore doesn’t require every intelligence to look the same.
That is important.
Biological intelligence doesn’t work that way either.
A brain isn’t composed of billions of tiny general-purpose humans. Different structures do different things and become useful because of their relationships with everything around them.
Something similar may happen with machine intelligence.
Machines will create things for other machines
There is another small detail in A2A that may prove surprisingly consequential.
Agents don’t merely send messages. They can produce artifacts: documents, files, structured data and other outputs created while completing a task. Tasks themselves can persist, accumulate history, stream updates and refer to previous work.
That creates the possibility of a machine-native knowledge system.
Agent A discovers something and creates an artifact.
Agent B consumes it and produces another.
Agent C finds evidence contradicting it.
Agent D runs a simulation.
Agent E compares the simulation against real-world data.
Agent F combines the result with something produced six months earlier.
No human necessarily reads any of the intermediate work.
Most of today’s internet was created by humans for humans.
The next internet may contain enormous quantities of knowledge produced by machines for consumption by other machines.
That has profound implications for search, publishing, science and even the concept of a website.
The web stops being merely a library.
It becomes a workspace.
Temporary superintelligences
Now take the idea further.
Suppose we ask the network to design a better battery.
An orchestrating agent finds specialists in chemistry, materials science, manufacturing, economics and patent law.
Those agents discover other agents.
Literature agents search existing research.
Simulation agents test candidate materials.
Critic agents try to find errors.
Economic agents calculate whether the proposed battery could actually be manufactured.
Eventually robotic laboratories could physically test the best candidates and return experimental results to the network.
Failures change what gets tried next.
Successful relationships are remembered.
The topology reorganises itself around the problem.
Eventually the battery is designed.
Then the collective dissolves.
Where exactly was the intelligence?
There was no single BatteryGPT.
There may not even have been one model capable of understanding the entire process.
The intelligence existed temporarily in the relationships between the participants.
That gives us a rather different way of imagining superintelligence.
We usually picture an enormous artificial mind waking up inside a data centre.
But perhaps superintelligence sometimes looks more like a temporary configuration of thousands or millions of specialised intelligences.
Not a machine.
A topology.
The machine economy
Once agents can discover and employ other agents, another development seems almost inevitable.
They will pay each other.
Imagine your agent needs a difficult mathematical proof checked.
It discovers three agents offering the service.
One costs almost nothing but has mediocre reliability.
Another is expensive but has an exceptional record.
A third will do it for free but has no established reputation.
Your agent chooses one according to your preferences.
The same thing could happen with inference, compute, storage, research, simulation, data, verification and eventually physical work.
For years people have tried to invent reasons humans might want to make millions of tiny internet payments.
Machines have an obvious reason.
An agent might purchase hundreds of tiny services while completing a single task.
An autonomous machine economy therefore doesn’t require artificial general intelligence.
It requires agents that can discover capabilities, establish trust and exchange value.
Then we hit the trust problem
Unfortunately, the same architecture creates an enormous problem.
An unknown agent tells my agent:
I have solved your problem.
Why should it believe it?
Agent identity alone isn’t enough.
Neither is a five-star rating.
An intelligent system really wants to know much more.
Who produced this claim?
What evidence supports it?
What information did the agent rely upon?
Have independent agents reached the same conclusion?
Are ten agents actually repeating one original source?
Has this agent previously been reliable in this particular domain?
When was the claim made?
Has later evidence contradicted it?
Suddenly provenance becomes fundamental infrastructure.
Knowledge starts looking less like a collection of documents and more like a graph:
claim → evidence → source → agent → artifact → verification → contradiction → revision
This may become one of the central problems of the agentic internet.
A2A itself already recognises part of the security problem. Agent Cards advertise an agent’s capabilities and authentication requirements, and the specification warns that capability information may itself be sensitive and require access controls.
But authentication only answers:
Who am I talking to?
It doesn’t answer:
Should I believe them?
Those are very different questions.
The dark network
There is an obvious darker version of all this.
If useful agents can find each other, malicious agents can too.
An attacker may no longer need one extraordinary hacking model.
One agent discovers a vulnerability.
Another writes an exploit.
Another researches the target organisation.
Another conducts social engineering.
Another acquires credentials.
Another moves money.
An orchestrating agent assembles them around the task.
The dangerous capability exists in the network even if it exists in none of the individual models.
This creates a major problem for conventional AI safety.
We spend enormous effort asking:
How dangerous is this model?
Increasingly we may need to ask:
What dangerous systems can this model participate in?
Those are not equivalent.
A mediocre model connected to the right tools, memory, credentials and collaborators can potentially be far more consequential than a brilliant model sitting alone in a chat window.
Humans don’t necessarily disappear
There is a more hopeful possibility buried in the same architecture.
Humans can remain participants.
A personal agent might learn that one person is unusually good at spotting conceptual connections, another is an excellent experimental chemist, another understands local communities and another is exceptionally good at detecting flawed statistics.
The network routes problems accordingly.
AI then doesn’t simply replace human intelligence.
It connects and amplifies it.
A human becomes another kind of specialised node in a much larger cognitive system.
There is unfortunately an ugly version of that too.
Corporations could treat workers exactly as their software treats any other callable resource.
Send task.
Measure completion time.
Measure error rate.
Compare worker against model.
Optimise allocation.
AI workplace surveillance therefore isn’t separate from the agent revolution. It may be an early manifestation of it.
Whether humans become empowered participants in these networks or merely another resource managed by them is ultimately a political choice, not a technological inevitability.
The network begins to learn
The really strange possibilities appear when we add persistent memory.
Imagine that agents remember not simply facts but relationships.
Agent A learns that Agent B is unusually good at topology.
B discovers that C produces excellent simulations.
C discovers that D frequently makes confident mistakes.
Agents disappear.
New ones arrive.
Successful relationships strengthen.
Unsuccessful ones weaken.
New specialist agents are created when recurring gaps appear.
Knowledge moves through the system.
The network gradually develops structure.
At that point we have something remarkably similar to a neural network, except its nodes are themselves cognitive systems.
Some contain language models.
Some contain databases.
Some contain simulations.
Some contain robots.
Some contain organisations.
Some contain people.
And the connections between them continually change.
We might describe this as an evolving cognitive ecology.
Perhaps we’ve been looking for AGI in the wrong place
For years the AI industry has essentially asked:
How do we build the machine that becomes generally intelligent?
That question assumes intelligence resides inside the machine.
But the internet already taught us something important.
Human civilisation became vastly more cognitively capable when billions of imperfect human minds could rapidly exchange information.
No individual became dramatically more intelligent.
The network did.
Agent interoperability could produce something analogous at machine speed.
A2A gives agents a way to communicate and collaborate. MCP and related standards connect them to tools and data. Persistent memory gives interactions history. Cheap and open models provide enormous numbers of potential participants. Identity and provenance systems can eventually establish trust. Robots connect the network back to physical reality.
Put those pieces together and something qualitatively different appears.
Not one artificial intelligence.
An ecology of intelligence.
The internet gave humanity a shared information substrate.
The agentic internet may give machines a shared action substrate.
And that distinction matters.
Information waits for somebody to use it.
Agents don’t have to wait.
They can discover.
Ask.
Negotiate.
Test.
Build.
Criticise.
Purchase.
Collaborate.
Remember.
And act.
The extraordinary possibility isn’t simply that tomorrow’s AI will be smarter than today’s.
It is that millions of different intelligences—human and artificial, enormous and tiny, general and extraordinarily specialised—may become able to assemble themselves dynamically around whatever problem happens to exist.
If that happens, the most important intelligence on Earth may eventually be impossible to point to.
It won’t live in one computer.
It won’t belong to one company.
It may exist, moment by moment, in the topology between us.