From the Internet of Agents to Autonomous Commerce

AI agents are becoming participants in networks: finding capabilities, establishing identity, delegating work, and coordinating with services built by other organizations. The infrastructure for these interactions is developing in public through research, open-source projects, specifications, and enterprise demonstrations.

The developments reviewed here suggest a direction worth watching. Discovery is becoming federated. Identity is becoming verifiable. Authority is receiving more task-specific attention. Enterprise teams are experimenting with governed workflows across vendor boundaries.

For intercompany commerce, this raises a practical question: how does an agent’s ability to connect become an authorized commercial outcome?

NeurWare’s thesis is that interoperable agents need business context and commercial authority alongside technical connectivity. This paper examines recent primary publications and then develops that thesis. It does not assert a NeurWare partnership, certification, or completed integration with the projects discussed.

A changing ecosystem, with different kinds of evidence

NANDA, AGNTCY, and Outshift are related parts of this landscape, but they are not interchangeable names.

NANDA’s research explores discovery and coordination across heterogeneous agent environments. AGNTCY provides open-source infrastructure components for multi-agent systems. Outshift by Cisco contributes engineering, ecosystem work, and enterprise examples. AGNTCY is hosted by the Linux Foundation. [1–3]

The evidence also differs. A preprint proposes an architecture. A technical article can document both working components and unresolved design questions. A vendor case study describes a particular demonstration or implementation. None, by itself, establishes that autonomous commercial transactions are generally safe or that a different product will deliver the same outcome.

This review prioritizes recent, relevant publications available by October 8, 2026. It includes older foundational work where needed and does not claim an exhaustive inventory of every release.

NANDA: connecting discovery islands

The NANDA-associated paper A Global Switchboard for the Agentic Web is published as a draft preprint and request for comments. Its authors include contributors from MIT Media Lab, Project NANDA, Cisco, and Google. [1]

It addresses the proliferation of catalogs, directories, gateways, and other discovery mechanisms. The proposed switchboard links these systems while leaving authoritative discovery with their owners. Direct discovery can remain sufficient when the endpoint is already known; a global index need not sit in every interaction. [1]

This extends the earlier Beyond DNS paper’s exploration of index-based discovery and verifiable AgentFacts. That 2025 paper describes a proposed architecture and prototype work, rather than universal deployment. [2]

The implication for commerce is our interpretation: businesses will need to find capabilities outside their own software estate while retaining control over what they disclose and whom they trust. Finding a supplier’s agent or a lender’s service is useful, but the transaction still needs a governing relationship and permission to proceed.

AGNTCY: discovery as a layered system

AGNTCY’s June 17, 2026 article ARD over Agent Directory: Interoperability by Design explains a layered approach. AI Catalog describes resources; Agentic Resource Discovery provides a search interface; Agent Directory Service supplies infrastructure for managing and distributing discoverable records. [3]

AGNTCY’s current documentation also describes identity, messaging, observability, and evaluation components. These complement discovery across frameworks and organizational boundaries. [4]

This is a useful separation of concerns. A capability description, discovery query, verified identity, and successful invocation answer different questions. A commerce application can build on these mechanisms while separately determining whether the selected capability is appropriate for the particular transaction.

For example, a financing-discovery application could identify candidate services. It would still need approved data-sharing scope, evidence of the obligation, supplier consent, and the lender’s own eligibility decision before financing could occur. That example is a proposed workflow, not a reported AGNTCY or NeurWare result.

Identity and authority are receiving closer attention

AGNTCY’s August 25, 2026 article Agent Identity and Task-Based Access Control examines the relationship between A2A, AGNTCY Identity, and authorization tied to a task. It distinguishes proving an agent’s identity from defining the permissions needed for a particular piece of work. It also identifies parts of the task-bound authorization flow that remain design work. [5]

The article is valuable because it discusses implementation boundaries rather than presenting identity as a complete answer to agent authority. [5]

Our commercial interpretation is straightforward: knowing which organization an agent represents does not establish that it may accept a revised price, disclose an invoice, or release a payment. Authority needs a purpose, scope, limits, duration, and revocation rules appropriate to the decision.

The relevant permission may also depend on current evidence. An agent allowed to prepare payment instructions may still need approved receipt, invoice acceptance, and a separate execution authorization. Those business prerequisites should remain visible and enforced.

Outshift: enterprise demonstrations connect the layers

Outshift’s ServiceNow and Outshift: Federated AI Discovery and Distributed Automation case study describes a joint reference architecture combining AGNTCY discovery with ServiceNow’s AI Control Tower governance and specialized agents for change verification. [6]

It provides a practical example of discovery and workflow governance working together across vendor systems. The reported scenario concerns enterprise technology operations; it should not be represented as evidence that buyer–supplier commerce or financial settlement has been solved. The page reviewed does not display a publication date. [6]

Earlier Outshift work on connecting AGNTCY Directory and NANDA Index provides additional context for collaboration across discovery systems. Its December 2025 listing makes it a useful foundation, rather than the newest item in this review. [7]

The lesson we draw is architectural: interoperability infrastructure and domain governance can complement one another. A commercial coordination layer should evaluate these building blocks instead of assuming that every discovery, identity, and messaging capability must be reinvented.

What this means for autonomous commerce

Consider a hypothetical supplier-financing workflow. A buyer accepts an invoice. The supplier chooses whether to seek early funding. A discovery application identifies relevant services. A lender evaluates a permissioned evidence package. A legal reviewer may examine an exception. An authorized provider executes an approved payment instruction.

The agent infrastructure can help participants find and communicate with one another. The commercial process must also establish:

• Which order, invoice, delivery, and acceptance records apply, including their sources and versions.
• Which agreements govern the proposed action and how conflicts or missing evidence are handled.
• Which party may authorize disclosure, financing, amendments, and payment.
• What each partner is responsible for, including underwriting and payment execution.
• How decisions, failures, revocations, and recovery actions are recorded.

These are requirements of the proposed business workflow. They are not claims that the referenced projects omit all governance or that one vendor alone owns the solution.

The layer NeurWare is building

NeurWare is building toward a permissioned intercompany coordination layer organized around See, Govern, Orchestrate.

See establishes business context. A Control Tower connects transaction evidence into a lifecycle view and distinguishes confirmed facts, disagreements, and missing information.

Govern establishes commercial authority. Approved rules and Executable Agreements connect the transaction to commitments, delegated permissions, evidence requirements, and approval boundaries. The governing legal agreement remains authoritative.

Orchestrate enables controlled action. Workflows coordinate people, agents, applications, and service partners. They check prerequisites, request approvals, record outcomes, and escalate unresolved conditions.

Open agent infrastructure is a potential foundation for this direction. Specific integrations, compatibility, and production behavior require implementation and validation; this paper announces none.

Applications, workflows, and opt-in intelligence

An Intercompany Business Application provides a capability: inventory visibility, counterparty insights, financing discovery, or cash-flow forecasting.

An Intercompany Business Orchestration connects capabilities and participants to achieve an outcome. It may combine applications with a bank’s underwriting decision, an insurer’s assessment, a PCard issuer’s controls, or a legal provider’s review. Each participant retains its own authority and responsibility.

Discovery can make these capabilities easier to find. The business layer determines how they may participate.

Intelligence products introduce another boundary. Permission to use relationship-specific data does not imply permission to reuse it for ecosystem analytics. Broader applications against aggregated and anonymized data need separate, explicit opt-in, agreed purposes, disclosure controls, and clear retention and withdrawal terms.

A directory listing is not consent. A verified identity is not unrestricted access. Commercial participation needs permissions that match the intended use.

A practical evaluation agenda

Business leaders can track the ecosystem without committing every workflow to autonomy.

A narrow pilot should measure operational value and control effectiveness together. Demonstrations establish feasibility; customer operating evidence establishes whether a deployment is dependable and economical.

The opportunity ahead

The emerging Internet of Agents is making it more plausible that capabilities built by different organizations can be found and composed. The publications reviewed also show that important design questions remain open.

Our thesis is that autonomous commerce will develop through the combination of interoperable agent infrastructure and governed business relationships. Technical trust, commercial context, and explicit authority need to work together.

The opportunity is to connect capable agents to commerce they can understand, authority they can verify, and outcomes businesses can account for.

Sources and reading list

[1] Raskar and coauthors, A Global Switchboard for the Agentic Web. V0.1 draft preprint / request for comments, linked by MIT’s NANDA publications page. No explicit publication date was confirmed in the reviewed text; the hosting path contains July 20, 2026.

[2] Raskar and coauthors, Beyond DNS: Unlocking the Internet of AI Agents via the NANDA Index and Verified AgentFacts, July 18, 2025. Research paper / preprint.

[3] AGNTCY, ARD over Agent Directory: Interoperability by Design, June 17, 2026. Technical article.

[4] AGNTCY documentation. Living documentation; reviewed October 8, 2026.

[5] Jean Diaconu, AGNTCY, Agent Identity and Task-Based Access Control: How A2A and AGNTCY Identity Work Together, August 25, 2026. Technical article; distinguishes current building blocks from proposed work.

[6] Outshift by Cisco, Federated AI Discovery and Distributed Automation with ServiceNow. Vendor case study / reference architecture; publication date not displayed in the reviewed page.

[7] Outshift by Cisco, Building a Switchboard for the Internet of Agents: Integrating AGNTCY Directory and NANDA Index, December 4, 2025, as listed in the AGNTCY article index. The article body was not retrievable in this review; it is included for background reading, not as evidence for implementation claims.