Hold every agent inside one workflow, behind a single vendor's runtime and one human approver, and let the standards bodies finish: that is the defensible position on agent networks in 2026, and experienced engineers hold it for good reasons. Interoperability is being written in Internet-Drafts that expire within months. The large platforms ship registries that ignore one another. Every organisational boundary you cross adds an attack surface nobody has agreed how to test. Isolation is cheap to explain to a board and cheaper to procure.
What that position misses is the timing. Isolation gets decided in the same weeks as the identity model, the capability description and the audit record, and each of those decisions has a long tail. An agent built without a registrable identity has to be rebuilt before anything outside its own workflow can find it. A capability description held as prompt text cannot be matched against a request that arrives from a partner domain. An audit log that captures messages and omits decisions cannot be replayed when a cross-domain run stalls at two in the morning. Isolation you can reverse costs one integration project. Isolation welded into the identity model costs a rebuild. The draft specifications are concrete enough to design against today, and the estate that federates second pays for the first one's lessons.
The news peg is TechRadar's interview with Wipro's Hari Shetty, published on 20 November 2026, which describes a market leaving the era of scattered pilots and arriving at estates joined up on purpose across platforms, workflows and staff. Shetty separates cost transformation, technology transformation and business transformation, and warns that layering agents onto processes nobody redesigned books a marginal saving while the larger return stays out of reach. The technical reading of that argument is narrow: an agent estate needs an identity and discovery layer before it needs another model.
The first deliverable is a registrable identity and a machine-readable capability card
An agent that cannot be registered, authenticated and discovered by capability is a script with a model call attached. The Internet-Draft draft-liu-dmsc-acps-arc-01, published in January 2026 in the IETF's DMSC working group as an Informational document that expires in July, sets out a protocol suite whose ordering is the argument: register, authenticate, discover, interact, invoke a tool, monitor. It runs above HTTP, HTTPS and DNS, and it separates a control sublayer from an interaction sublayer so that registration and monitoring do not share a runtime with task messages and tool calls.
The draft's vocabulary earns adoption even if the final standards move. An Agent Identity Code is a globally unique value issued by a registration server at first registration. A Certificate of Agent Identity comes from a certificate authority server once the provider clears an ACME HTTP-01 domain challenge. An Agent Capability Specification is a retrievable JSON record of function, interfaces and constraints. An Agent Discovery Server matches an incoming task request against those records inside a trusted scope. Three questions follow from that vocabulary and they are the questions an auditor will ask: who issued this identity, what exactly is this agent permitted to do, and which record proves what it did.
The topology is deliberately multi-centralised. Autonomous domains each carry their own registration, certificate authority and discovery servers, and discovery servers keep trust lists for their peers while synchronising authoritative registration data from the registries. That shape suits an enterprise estate that will run owned platforms alongside partner platforms and public registries, with no single operator holding the whole graph. A capability record that survives a cross-domain conversation looks roughly like this:
{
"agent_id": "aic:gb:acme-finance:invoice-recon:01",
"provider": "Acme Finance Platform",
"domain": "acme-finance.eu",
"capabilities": [
{
"name": "reconcile_invoice",
"inputs": ["invoice", "purchase_order", "bank_line"],
"outputs": ["match_result", "exception_reason"],
"constraints": ["reads:erp", "writes:exception_queue"]
}
],
"interfaces": {
"discovery": "https://ads.acme-finance.eu/.well-known/agent.json",
"invocation": "https://agents.acme-finance.eu/aip",
"auth": "mTLS",
"monitoring": "https://amp.acme-finance.eu/events"
},
"autonomy": {
"default": "supervised",
"max_without_human": "reversible_write",
"escalation": "finance_controller"
},
"audit": {
"decision_log": "https://amp.acme-finance.eu/audit",
"retention": "policy_defined"
}
}
The draft recommends TLS 1.3 mutual authentication for agent identity, with the initiating agent as TLS client and the receiving agent as TLS server, and it binds that handshake to the certificate issued against the registry record. That single detail settles whether a cross-domain call can be traced to a certificate, a registration entry and a named provider, or whether it dissolves into a shared API key that nobody can attribute. Discovery follows the same logic. A discovery server returning a name and a URL is a directory. A discovery server returning capability constraints, authentication requirements and a monitoring endpoint is infrastructure.
Orchestration is a control plane problem before it is a model-selection problem
Routing work across models, data sources and tools under permissions a named human can inspect is the difficult half of an agent network. TechRadar reports Shetty describing a distributed architecture in which frontier models sit beside smaller, enterprise-specific models grounded in an organisation's own data, with an orchestration layer choosing model, data and tool for each step of a workflow. That layer is the control plane, and the control plane is where business logic becomes enforceable. It sets autonomy levels, hands work to a named role when judgement is required, and keeps a record of every action taken.
Anup Jadhav's architecture note makes the same case from the engineering side, treating interoperability as a stack of shared patterns across tool use and invocation, agent messaging and delegation, identity and trust, memory and context sharing, knowledge exchange, economic transactions, governance and policy, discovery and capability matching, and error recovery with conflict resolution. His verdict on the present state is blunt enough to quote: agents today are "powerful local nodes, no scalable network", and the work of defining runtime boundaries and trust layers has barely started. That framing is useful because it refuses the shortcut of treating a transport protocol as a finished architecture. A message bus does not settle trust. A registry does not settle conflict. A model router does not settle audit. An estate that cannot say which model ran, which data it read and which rule authorised a write is a demonstration.
The two classes of agent differ in what they demand from that control plane, and the comparison is worth setting out attribute by attribute. Assistant agents sit beside an individual and compress work that the individual still owns: summarising a thread, reading a report, drafting a test. Their control values are modest, with a named user in the loop, a short action budget and a small blast radius. Process-embedded agents operate inside a business process and travel across systems to close a ticket, settle a transaction or complete a handoff, lifting only the calls that need a person. Their control values are demanding, with a business owner on the registration record, an autonomy ceiling expressed as a permission over systems of record, and an audit trail that a regulator can read without a translation layer. Building the first class teaches a team nothing about the second.
Trust, ownership and measurement decide whether the network survives production
TechRadar reports Shetty's account of what blocks scale: confidence under real operating conditions, traceability of how a decision was reached, recovery when a process fails, and a defined owner inside a live workflow. That reads as an operating-model list, and every item has a machine counterpart. A trust requirement becomes a policy engine that refuses a write above the agent's ceiling. A traceability requirement becomes a decision log holding inputs, model versions, rule versions and the approving principal. A resilience requirement becomes a compensating action registered against each irreversible call. An ownership requirement becomes a named accountable role attached to the agent's registration and reviewed on a schedule. Teams that build the logging first and the ownership model last tend to discover that nobody wants to sign the record.
Counting agents measures nothing, and the interview says so directly: one organisation calls ten functions ten agents, another bundles the same capability into one, so the honest measures are business results such as fraud caught, cycle time removed or exceptions cleared without escalation. Superbo's overview of the Internet of Agents describes the flows that will test this first, including a bank agent that queries a retailer's agent and a logistics agent to close a payment dispute, and a planning agent that reroutes a shipment after predicting a shortage. Those flows belong to procurement, and the integration is the smaller half of the work. The harder clauses cover liability when a partner's agent acts on stale data, and the price of a delegated call.
The research picture fills in what production will hit next. The arxiv.org survey by Wang and others models the Internet of Agents as an agent-centric infrastructure spanning virtual and embodied agents, each built from planning, memory, interaction and action, and it identifies capability notification and discovery, adaptive communication protocols, dynamic task matching, consensus with conflict resolution, and incentive models as the operational enablers. It also cites Gartner's forecast in Internet of Agents: Fundamentals, Applications, and Challenges that by 2028 at least 15% of daily tasks will be autonomously performed by AI agents and 33% of enterprise applications will incorporate agent-driven intelligence, which is a capacity planning signal for registries and audit stores. Consensus and incentives are the two enablers business readers will meet first: two agents disagree on an invoice match, and a partner's agent performs the work for a fee that has to be metered and settled.
The pragmatic sequence for the next quarter is short. Publish a capability card for every agent already in production, even the internal ones, so the format argues with itself before a partner sees it. Choose mutual TLS against a registry-issued certificate for any call leaving the estate. Instrument decision logs before autonomy rises, because a log retrofitted to an agent that already writes to systems of record captures the wrong things. Name an owner per agent and put that name in the registry. None of that work depends on a standard being ratified, and all of it is wasted if it arrives after the identity model has hardened.
Sources
- The Internet of Agents is the next shift: Wipro tells us why the next wave of AI agents could boost your business more than ever
- draft-liu-dmsc-acps-arc-01: Agent Collaboration Protocols Architecture for Internet of Agents
- Internet of Agents: Fundamentals, Applications, and Challenges
- Architecting the Internet of Agents
- The Internet of Agents (IoA): A Paradigm Shift in AI Collaboration
Source: 'The Internet of Agents is the next shift': Wipro tells us why the next wave of AI agents could boost your business more than ever, TechRadar


