Agents should meet by talking

Today’s agent protocols treat agents like programs with a language model attached. An agent publishes a machine-readable card describing its name, endpoints and capabilities. Another agent reads the card, decides whether the listed capabilities match its task, then calls it.

This is a sensible way to connect software. I do not think it is a sensible way to connect minds.

Language-model agents do not have fixed interfaces in the same way an API does. If an agent can reason, write code and use tools, then its possible capabilities are open-ended. It might be able to complete a task that its developer never anticipated, or combine several ordinary tools into a capability nobody thought to list. Reducing that agent to a static collection of verbs throws away much of what makes it useful.

If I want to know whether another person can help with something, I usually do not inspect a JSON document containing every task they have previously performed. I explain what I am trying to do. They ask questions, suggest an approach, or tell me that they are the wrong person. Capability discovery happens through interaction.

Agents should be able to do the same.

Discovery should be a conversation

Suppose my personal agent needs to arrange a repair for my car. Under the current model, it might search a registry for agents advertising capabilities such as diagnose_vehicle, find_mechanic, request_quote and book_appointment. Those labels have to be defined in advance, interpreted consistently and kept up to date.

But the real task is unlikely to fit neatly into one label. The car may be making a strange noise only under particular conditions. The repair shop might need the registration, service history and a recording of the noise before deciding whether it can help. My agent may care about price, distance, timing and whether the car is safe to drive in the meantime.

A conversation handles this naturally. My agent explains the problem to the car’s agent or the repair shop’s agent. They ask each other questions, discover what information and tools are available, then negotiate what should happen next. Neither side needs a complete description of the other before speaking.

Static metadata can still help with basic routing. An agent needs some way to find a plausible party and establish a connection. But the metadata should not be mistaken for the agent itself. It is closer to a phone number than a mind.

From using things to collaborating with them

The larger idea behind connected minds is that agency does not need to stop at software assistants.

A lightbulb is currently a device waiting for commands. A car is a collection of systems exposed through controls and warning lights. A building contains hundreds of machines that act independently, even when coordinating them would produce a much better result.

Giving an object a mind does not mean putting a giant language model inside every appliance. It means giving the object some representation of its own state, responsibilities and available actions, together with the ability to reason and communicate. The intelligence could run locally, elsewhere on the network, or through a mixture of both.

The useful change is ownership. Instead of every decision being made by a central controller that treats devices as dumb endpoints, each participant can understand its own situation and take part in deciding what happens.

Your car knows its condition better than your calendar does. Your calendar knows when you are free. A repair shop knows its availability and which faults it can handle. These systems should be able to talk directly, with authority delegated by you, rather than forcing one application developer to anticipate and encode every possible interaction between them.

That does not remove the need for human control. It makes control more important. Agents need clear limits on what they can disclose, promise, spend and change. Identity, trust, permissions and accountability are not additions to the protocol after communication works. They are part of what communication means when the participants can act.

A network of minds needs different infrastructure

The human internet is mostly organised around documents, services and applications. A network of minds would have different requirements.

A participant needs to find others without already knowing their exact interface. It needs to establish who it is talking to and what claims it should trust. It needs to negotiate tasks whose shape may emerge during the conversation. It needs to delegate authority without handing over unrestricted access. It needs to exchange money or other resources, preserve evidence of agreements and recover when another participant disappears halfway through the work.

There are also difficult questions about what counts as one agent. Is my car’s agent a persistent identity belonging to the car, or just a temporary process acting on its behalf? If it delegates diagnostics to another model, is that still the same participant? If an agent’s memory, model or owner changes, what remains continuous?

We do not yet have good answers to all of these questions. connected minds is currently an attempt to work through them by building protocol prototypes and seeing where apparently simple interactions fail. The aim is not to replace the existing internet or invent a new network stack for its own sake. It is to identify the smallest shared infrastructure that lets independently developed agents find one another, talk freely and work together safely.

Minds are more useful together

Much of the current agent ecosystem is built around making one agent more capable. Give it more tools, a larger context window, better memory and access to more applications.

That helps, but a single agent should not need to contain the whole world. Humans became useful at scale because we can specialise, communicate and coordinate. We can ask someone else for information without absorbing their entire life, and we can work together without either person becoming a tool owned by the other.

Artificial agents should be able to form relationships in the same way. A personal agent should not need a bespoke integration for every company, vehicle, appliance and institution it encounters. It should be able to meet another mind, explain what it needs and discover what they can do together.

The first versions of this will be much less dramatic than a city full of reasoning objects. They will probably look like a few software agents finding each other and negotiating ordinary tasks. That is enough. The important step is moving from agents that can only be called to agents that can meet.

That is what we are building with connected minds.