What is the Internet of Intelligence (IOI)?

What is the Internet of Intelligence?

The Internet of Intelligence (IOI) is what exists when autonomous software owned by different parties can do real, consequential work for one another — safely, verifiably, and for value — and the open operating stack that makes it possible. Not conversation, not content: work. Its systems act only under bounded authority, and they produce a verifiable record of what they actually did. "Open" means the protocols and architecture are published, so no single vendor owns the boundary.

For example: a support system that can issue refunds up to its declared limit, a data agent that can read the CRM but never the payroll database, or a drone fleet that accepts survey missions but cannot leave its approved operating area — each leaving a record that can be audited, disputed, or accepted. You declare what a system may do, run it where you choose, and get back a signed record of everything it did.

How it works

AI systems can already reason, write code, and call tools. What is missing is a trustworthy action layer for software that touches money, infrastructure, contracts, and production systems. Without one, a user has to trust a model provider, an agent host, and a human operator all at once.

An Actor is autonomous software that can take consequential action, such as spending money or deploying code, on behalf of a person, organization, system, or another Actor. A Worker is the role an Actor takes when it is hired for bounded labor: it receives scoped authority, performs work, and produces receipts, the signed records of what it did. A model supplies cognition. An Actor supplies agency. A Worker supplies bounded, accountable labor.

The stack turns intent into a transaction-shaped workflow:

1. A person, business, or software system states an outcome.
2. An Actor or group of Actors receives scoped authority.
3. The Actor executes through models, tools, APIs, devices, compute, services, or workflows.
4. The Hypervisor runtime checks each proposed action against deterministic rules before it takes effect.
5. The runtime records receipts, state transitions, and evidence.
6. The result can be inspected, replayed, disputed, paid, or composed into later work.

Step 4 is the Determinism Boundary. Above it, models plan, infer, search, rank, and draft non-deterministically. Below it, any consequential action must be canonicalized, authorized, logged, and receipt-bearing. The goal is not to make model thought predictable, but to make real-world consequences bounded, inspectable, replayable, and enforceable.

Models can remain probabilistic. Consequences cannot.

Work, not weights

The Internet of Intelligence is not a shared or global model, not a federated-learning network, not an agent swarm, and not one provider's agent ecosystem. It exchanges bounded, attributable, challengeable work — not weights.

The deepest rival reading comes from federated learning, where parties cooperate by pooling knowledge and the learning protocol is the coordination substrate. That is coherent inside one trust domain, such as a lab consortium. Among self-interested, independently governed institutions, it fails for four reasons.

It was built against the wrong adversary. Federated learning defends against honest-but-curious participants. Once participation carries economic stakes, the adversary becomes economic: sybils, free-riders, poisoning, work claimed but never performed. Adversarial contribution has to be attributed and made accountable per contributor, not averaged away.

It is silent at the effect boundary. As a training-time protocol, it has nothing to say once machine work has effects in other parties' systems. There, either delegation is bounded, revocable, and receipted, or the network oscillates between paralysis, where nobody delegates, and ambient authority, where everybody regrets it.

Weights launder rights. A gradient or parameter delta that crosses an ownership boundary cannot be attributed, revoked, or licensed afterward. That is why cross-competitor federated learning has effectively zero production footprint after a decade.

There is no incentive layer. No institution strengthens a shared model its rivals also use, uncompensated. Sustained cooperation among strangers requires pricing, attribution, dispute, and settlement. A labor market, not a potluck.

The two models are mirror images. The pooling model assumes cooperation and engineers privacy: it describes how a lab consortium behaves. The Internet of Intelligence assumes sovereignty and engineers cooperation: it describes how institutions behave.

Humanity has built an internet of intelligence before, for human intelligences. Its substrate was never pooled cognition. It was identity, contracts, agency law, attribution, courts, reputation, and money. The machine version needs machine-speed versions of the same primitives:

• Leases, scoped and revocable grants of authority, for agency law
• Receipts for attribution
• Challenges for courts
• Sparse settlement for money

Where the analogy runs out. Human institutions can compel; the machine substrate cannot, and a counterparty who has taken a model weight across the boundary cannot be made to un-know it. These primitives reproduce agency, attribution, adjudication, and payment, not the sheriff.

None of this refutes federated learning. It demotes it: cross-boundary training and distributed inference become governed workloads that run on top of the substrate, not the substrate itself.

Other definitions of the Internet of Intelligence

The utility reading (arXiv:2205.08977, arXiv:1909.08068) promises to make intelligence "as easy to obtain as information." Model APIs delivered that, and the result was cognition everywhere, accountability nowhere. The capacity reading, from 6G research, is about carriage; the telecom question in the FAQ below returns to it. The connectivity reading (arXiv:2407.07061, arXiv:2505.21550) standardizes how agents talk, not what they may do or on whose authority. Each names a genuine ingredient; none, done alone, produces an internet of intelligence.

Where verification is cheap enough

Verifier economics, not authority plumbing, is the binding constraint. Much intelligent work is nearly as expensive to verify as it is to produce, and pricing unverifiable work silently reintroduces the trust the substrate exists to remove. A receipt proves what happened. It does not prove the work was any good.

The substrate works now where verification is structurally cheaper than production:

• Code with tests: running the suite costs a fraction of writing the code.
• Structured output: schema validity, type checking, and referential integrity are mechanical.
• Bounded transactions: a purchase is checkable against its declared vendor, price, budget, and jurisdiction.
• Reproducible operations: a deployment replayed against a recorded state root reproduces it or does not.

It works poorly for open-ended judgment: strategy, research direction, design taste, most writing. Adding receipts to a bad essay produces a well-documented bad essay. There the network falls back on what human markets use: reputation, repeated dealing, and staged exposure.

What would prove it

The Internet of Intelligence has been demonstrated when an independently operated, external Worker can discover eligible work, negotiate terms, receive bounded leases, do the work, return a verifiable contribution with its credit and dispute lineage intact, and leave with a portable bundle of its own state, without needing to share a runtime, a database, or an administrator with anyone, or to keep trusting anything IOI hosts.

No implementation has passed this test, including this one; the remaining distance is tracked in public on the roadmap.

The real rival

The real rival is not the pooling model. It is ambient-authority convenience: an agent handed broad standing access because that is easier, and everyone hopes. Agent platforms are its near form; the IOI stack is the authority and settlement layer beneath them, not a competitor.

Convenience does not lose everywhere. The sovereign version is likely to win where work carries real liability: regulated industries, cross-company transactions, infrastructure with blast radius, anything an auditor or a court may ask about. There, somebody is accountable by name, and "the model decided" is not an answer they can give. It may never win consumer agents: a personal assistant booking dinner inside one vendor's account has no counterparty, no dispute, and no liability worth the overhead.

Which parts are claims and which are proposals

The authority half is a claim about the world. Wherever non-human actors take consequential action across ownership boundaries, delegation must be scoped, revocable, and receipted, or the network falls into paralysis or ambient authority. That holds under every market structure, including one dominated by a few model providers.

The economic half, an open labor market, is a bet about how an industry develops, not a theorem.

Everything specific is a proposal: the Hypervisor, leases, receipt graphs, the interop protocol, and sparse settlement are one worked answer. "Cryptographic" is the implementation, not the point: the necessity attaches to properties such as tamper evidence and revocability, not to any particular chain.

Why not assemble this from what already exists? Identity providers, enclaves, signed logs, and payment rails each stop one step short: an enclave attests that a computation ran as built, not that it was permitted. Wire them together and every organization still builds its own authority layer: once, badly, and unportably, in a private vocabulary that a counterparty cannot evaluate. The stack is built to be that missing layer, over those components rather than against them: leases, scopes, revocations, and receipts that bind an action to the authority that permitted it and the evidence it produced. The argument does not by itself license a token or mandatory public settlement.

IOI in five parts

1. Hypervisor and Determinism Boundary: the runtime that checks proposed actions against deterministic rules and emits receipts.
2. Actors and Workers: each Worker declares a role, policy envelope, receipt obligations, and settlement posture.
3. Routing and contracting markets: not a directory of models, but a way to bind work to evidence, resolve disputes, and settle payment after verified delivery.
4. Edge-in state, memory, and private workspaces: state stays close to where work happens, and confidential compute supplies context without giving untrusted hosts plaintext custody of data or credentials.
5. Sparse settlement and arbitration: public settlement only where it adds trust, such as escrow, disputes, and final settlement.

Core concepts

The machine economy is the setting: autonomous software owned by different parties doing paid, consequential work for one another.

Receipt-carrying agency means every consequential action produces a signed record. Multi-worker workflows form receipt graphs that make labor auditable across parties.

Mixture of Workers is the market architecture for routing accountable labor across specialized Workers without requiring any Worker to reveal its private model, prompt chain, or proprietary method.

Examples of Internet of Intelligence work

• An Actor might fix a software issue and produce a receipt linking the prompt, code diff, tests, approvals, and deployment boundary.
• A Worker might buy inventory only if vendor, price, budget, jurisdiction, and approval policy match a published manifest.
• A verifier Worker might check another Worker's output against an objective test or schema, and its receipt decides whether payment settles.

Architecture links

Each layer of the IOI stack has one home: Hypervisor for execution, with the Hypervisor Daemon underneath it; the agents layer for agents and agent systems; and the services layer for Service-as-a-Software. The IOI Roadmap links them to implementation records and evidence.

Developer resources live in the public ioi repository, with protocol reference material at docs.ioi.network and source and release evidence in the IOI Foundation GitHub organization.

Papers and evidence

• IOI Technical Whitepaper: protocol architecture, deterministic machine authority, receipt-carrying agency, Worker routing, and settlement.
• Mixture of Workers: the accountable labor market architecture.
• Alignment Security and the Boundary of Machine Authority: the authority-bound safety thesis.
• Deterministic Authority for Machine Action: the valid-action boundary around mutable machine actors.

About IOI Foundation

IOI Foundation's work centers on the machine economy, verifiable machine labor, deterministic authority, receipt-carrying agency, and settlement infrastructure for autonomous software Actors.

FAQ

What is the Internet of Intelligence?

The Internet of Intelligence (IOI) is the network condition in which independently governed systems exchange bounded, verifiable work under declared authority — an economy of accountable machine labor rather than a pool of shared models. IOI is the open operating stack for autonomous systems that do real work under real limits: AI agents and autonomous software that can act — spend, provision, file, operate — only under bounded authority, and that produce a verifiable record of what they actually did. One governed system can run across many machines, sites, people, and devices while behaving as one accountable system with one history. Independently owned systems can also work together through AIIP, the interop protocol between autonomous systems — but only under terms both sides accept. IOI L1, an optional shared trust layer, exists for the few commitments that need public settlement; running an IOI system does not require it.

What does IOI stand for?

IOI stands for Internet of Intelligence: the open operating stack, published by IOI Foundation, for building and running autonomous systems that act within declared limits and produce evidence of what they did.

Is IOI shared or federated intelligence?

No. The Internet of Intelligence does not pool intelligence. Independently owned systems keep their models, data, and methods private — and still do accountable work for each other. What crosses the boundary is an agreement, the authority it grants, and proof of what was done. That is a cryptographic labor economy: machine work bound to declared authority, evidenced by receipts, and settled under terms both sides accepted. IOI's architecture uses a local-first, fractal blockchain topology so authority, rights, and evidence remain portable and independently verifiable across trust boundaries. Work stays at the edge; shared settlement is invoked only where independent rights, dispute resolution, reputation, or economic finality require a neutral trust root.

Is the Internet of Intelligence a telecom or 6G concept?

No. In network research, "internet of intelligence" sometimes names AI embedded in 6G-era telecom infrastructure, and elsewhere it can mean pooled model training or generic agent networks. Those readings optimize capacity — moving data faster, training better shared models. Capacity is not what is missing. The term names something wider than a network upgrade: the condition in which independently governed systems exchange bounded, verifiable work under declared authority. Faster networks and cooperative model training are workloads that run on such a substrate, not substitutes for it. IOI is built to implement that condition, and it is independent of any radio or telecom architecture.

What is a bounded autonomous system?

A bounded autonomous system is software that pursues goals and takes consequential action inside an explicit constitution: a declared rulebook covering who can authorize what, which policies apply, who its members are, what evidence it must produce, and how it recovers when something fails. A bounded system is useful entirely on its own, and its owner can keep it fully independent.

Can one IOI system run across multiple machines or sites?

Yes. One governed system can place work across execution nodes, verifiers, gateways, people, sensors, robots, and local controllers while keeping one identity, one authority model, and one record of what happened. Coordinating the members of one system is built into the stack; it does not require the cross-organization protocol.

When does AIIP apply?

AIIP is the interop protocol between separately owned autonomous systems, and it applies only at that boundary: two independently governed systems exchanging bounded work, evidence, and contribution under terms both sides accepted. Being discoverable or invited creates no authority and no obligation.

Is IOI L1 required?

No. IOI L1 is optional. It is a shared trust layer for commitments that genuinely need public settlement — selected registry, rights, assurance, or dispute needs — and it is offered only where demonstrated demand justifies it. An IOI system operates fully without it.

What does an enterprise retain with IOI?

Your ontology, approved memory, corrections, private evaluations, workflows, policies, evidence, and lineage. They stay inside your governed boundary and remain portable across model and provider changes, so models stay replaceable suppliers rather than the owners of your operational knowledge.

What is a Worker?

A Worker is the bounded labor actor inside an IOI system: it receives scoped authority, performs work, and produces receipts. The model behind a Worker is replaceable cognition supply — authority and accountability attach to the Worker, not the model.

What is verifiable machine labor?

Verifiable machine labor is autonomous work backed by evidence: receipts, approvals, state changes, and acceptance records that can be inspected, replayed, disputed, paid, or composed into later work. It is a primary application of the IOI stack, not the definition of the whole system.

What is machine authority?

Machine authority is the declared boundary for consequential action: what a system may do, which systems it may touch, how much it may spend, and what evidence must exist afterward. It is granted by a person or organization, which stays accountable, and IOI makes it enforceable through explicit authority, policy, receipts, and optional shared settlement.

Who is building IOI?

IOI is published by IOI Foundation through internetofintelligence.com, the official canonical domain for the Internet of Intelligence.