The Task-Driven Agent Internet: Re-framing the Network Role and the Critical Path

The Reconstruction of the Network’s Role and the Critical Path — As the Internet’s service object shifts from “human clicks” to “task execution,” the network is undergoing a role reconstruction from “bit pipe” to “task nervous system.” Based on the latest wide-area network traffic measurement data, combined with global standards and industry progress, this report answers a core question: what role should the network play in the agent Internet, and what is most urgently needed right now.

Executive Summary

1. The paradigm shift is established. The Internet is moving from “connecting information” and “connecting people” toward “connecting intelligence and tasks.” The metric of scale is shifting from MAU (monthly active users) to DAA (daily active agents) — what users pay for is “the job gets done,” not “whether AI can do it.” I

2. The character of traffic has changed. The network traffic triggered by an agent executing the same task is about 450% more than manual operation F, of which about 70% is model inference traffic F; inference flows “last longer (about 2x) and run at lower rates (about 1/10),” with uplink share anomalously rising to 9% F. Network pressure is shifting from “peak bandwidth” to “state management and uplink bearing.”

3. The network’s role needs reconstruction. The AI inference path is becoming a “strategic network asset” F. The network will upgrade from a “pipeline carrying bits” to an “execution foundation carrying tasks,” and its core value anchor can be stated in one sentence: the determinism and trustworthiness of tasks. I

4. Three bottleneck links. Unified agent identity and capability discovery, cross-domain trusted identity and authorization, and task-level quality of service assurance — these three are globally recognized gaps and also where value is most concentrated. I

Data annotation system: F fact (with traceable source) · I inference or forecast · A assumption · E estimate.

Core Conclusions

The task-driven agent Internet is the direction of next-generation network evolution; the highest priority on the network side is not capacity expansion, but completing the three foundational capabilities of “identity—discovery—assurance.”

Supporting arguments

1. Demand side has already scaled: a single consumer health agent reached 30 million MAU F, the median agent deployment among leading enterprise organizations has reached 23 F, and globally active agents are projected to reach 2.216 billion by 2030 I. Scale is no longer the question mark; the bottleneck is “reliable delivery.”
2. The gap on the network side is already recognized: in IETF’s latest survey of agent discovery mechanisms, six major unsolved problems are explicitly listed (capability discovery, federated interoperability, connecting the network layer with the application layer, etc.) F, all pointing to the network layer.
3. Technical paths are taking shape: network-layer mechanisms such as semantic routing, computing-power routing, and in-network state awareness are being proposed, and can directly respond to the new characteristics of “long-lived inference flows, strong state, heavy uplink.” I

Kill Conditions

• If mainstream agent architectures shift comprehensively toward on-device standalone closed loops (inference does not leave the device, and agents almost never communicate across domains), the necessity of “the network carrying tasks” will be greatly weakened.

• If application-layer protocols (MCP / A2A / ATH, etc.) fill in discovery, trust, and quality of service assurance on their own, and explicitly do not require network-side participation, then the room for the network layer to maneuver will narrow significantly.

I. Origin: The Internet’s Third Leap

To judge a research direction, one must first see whether it is stepping on the beat of a paradigm shift. Regarding the agent Internet, industry has formed a rare consensus judgment — the Internet is undergoing its third leap.

First generation

Connecting information

The subject is people, addressing relies on domain names and IP, and the bottleneck is bandwidth and latency. People actively search web pages.

Second generation

Connecting people · services

The subject is people plus devices, addressing relies on services and accounts, and the bottleneck is connection count and experience. People click apps to issue commands.

Third generation

Connecting intelligence · tasks

The subject is people plus agents, addressing relies on capabilities and intent, and the bottleneck is task completion rate and trustworthiness. The goal is goal-driven, automatic networking.

CAICT summarizes the core of the third-generation Internet as a leap “from connecting people and connecting information toward connecting intelligence, connecting decisions, and connecting tasks” F. This means the basic unit of network interaction has changed from “bytes” to “tasks.”

For this reason, “task-driven” is more accurate than “task-oriented service.” “Task” and “service” belong to two different levels — traditional service-oriented architecture (SOA) emphasizes capability encapsulation, whereas the agent Internet emphasizes the closed loop of task execution. Using “task-driven” as the modifier both distinguishes it from the traditional Internet (connection-oriented, location-oriented) and avoids conceptual confusion with service architecture.

At the terminology level, “Task-oriented Communication” is already a recognized hot term in 6G research, so “task-driven” has an international terminological basis and carries no risk of being a coined phrase.

II. The Question That Must Be Answered: What Has Agent Traffic Changed

To judge what the network should do, one must first look at what agent traffic actually “looks like.” A recent wide-area network report based on real network measurements explains this change more thoroughly than traffic growth itself. Its value can be summarized in three progressive layers.

Layer One · Traffic Has Not Just Increased, Its “Character” Has Changed

In one public information-gathering test, an autonomous agent, to complete a single task, autonomously communicated repeatedly with web tools and models, reaching 44 web sources in total. The result — for the same matter, the network traffic triggered by the agent was about 450% more than manual operation F. The measurers called the connection between the agent and the model the “Spinal Cord”: this is the new critical path, and once it jitters, the task chain fails directly, not merely a page slowing down.

One user intent = dozens of machine-to-machine network calls. To the user it is “one thing”; to the network it is “a pile of flows,” and these flows are not equal in importance.

Layer Two · It Is Like a “Thin Water Pipe,” Not a “Large Water Pipe”

Network traffic increment

+450%

Agent completing similar tasks vs manual

Composition of new traffic

about 70%

is model inference traffic

Duration of a single flow

about 2x

inference flow vs ordinary web transaction

Median flow rate

about 1/10

inference flow vs ordinary web transaction

Flows where uplink exceeds downlink

9%

ordinary HTTP only about 0.5%

Inference traffic growth rate

about 10x/year

some networks measured about 4x in 8 months

A large model generates one token at a time, so inference flows exhibit the counterintuitive characteristics of lasting longer and running at lower rates. This means the pressure is not mainly on peak bandwidth, but on state management — devices such as firewalls, intrusion detection, and deep packet inspection that need to maintain flow state must all re-plan table capacity for “long connections × massive flows.”

Layer Three · Uplink Traffic Has Been Reversed

In inference flows, about 9% have uplink exceeding downlink, whereas in ordinary HTTP transactions this proportion is only about 0.5% F. The reason is that prompts have evolved into “context packages”: historical conversations, file summaries, web content, tool-call results, task state — all must be pushed upstream. Uplink capacity, wireless-side planning, enterprise egress, and security policies all need to be re-examined accordingly.

On Latency: The Network Is Not the Main Bottleneck Now, But It Is Approaching the Main Battlefield

Current end-to-end inference latency is mainly determined by model processing time, with network latency accounting for only a smaller part. But as inference hardware continues to accelerate and model processing time falls, the share of network latency in the overall experience will rise — the network “is not yet the main bottleneck, but will get closer and closer to the main battlefield.” The report’s long-term forecasts (such as the share of inference traffic in 2035 and multiple-fold growth in consumer and enterprise traffic) should be treated as planning signals rather than established facts I.

The Real Question: Should the Network “Understand” Agents

The report’s landing point is: the AI inference path will become a strategic network asset F. In the future, enterprises cannot only ask “is the network connected,” but also “is the critical AI path stable, observable, and assured.”

But it must be pointed out: this kind of discussion only raises the question; it has not yet given the network-layer answer. It explains clearly what the network will encounter, but does not answer “how to identify, how to assure, and at what granularity to schedule.” This vacuum is precisely the value zone of network-layer research — and also the part this report will fill in later.

III. Global Research Landscape: Who Is Doing What

By organizing global progress into a map by “participant—progress,” the division of competition and cooperation becomes clearly visible.

Level	Main drivers	Key progress
Conceptual paradigm	CAICT, major telecom research institutions	Three-generation theory: connecting information → connecting people / services → connecting intelligence / tasks
Network architecture	Operator consortia, equipment vendors	Three-layer (bearer-service-application) or four-layer architecture + task-oriented dynamic networking; AI-native architecture adds an intelligent communication layer + semantic intent layer
Application-layer protocols	Anthropic, Google, Linux Foundation	MCP (agent↔tool), A2A 1.0 GA (agent↔agent), ANP, etc.
Trusted protocols	CAICT together with the three major operators and vendors	ATH 1.0 (user-agent-application three-party nine-step handshake), national standards for agent interconnection
International standards	IETF / 3GPP / ITU-T / ETSI / TM Forum	IETF CATALIST BoF; 3GPP TR 22.870 (agents as first-class entities in 6G SBA); ITU-T SG13 AI-native network research group; ETSI ENI / ZSM; TM Forum L5 autonomous networks
Key enablers	Academia and the network side	Semantic routing, computing-power routing, capability discovery, task orchestration, DID / VC identity system

Most Worth Watching: Six Recognized Gaps

In IETF’s latest survey of agent discovery mechanisms, the authors explicitly list six problems for which no standard solution yet exists F:

1. No standard for capability discovery — agents cannot be retrieved by “capability” alone, only by knowing the address in advance.
2. No interoperable federation — directories of different implementations cannot exchange registration information, and trust models are incompatible with one another.
3. The network layer and application layer are not connected — agents discovered at the network layer cannot be mutually recognized with A2A agent cards and MCP tool schemas.
4. Fragmented metadata schemas — each party defines its own format, requiring customized translation between them.
5. Difficult trust bootstrapping — “found” does not equal “trusted,” and root trust remains unsolved.
6. No lifecycle management — capability changes cannot notify clients that have already discovered it.

These six items are essentially all network-layer problems. Together they point to one conclusion: what the agent Internet lacks most right now is not smarter models, but a public foundation that lets agents “find each other, recognize each other, and connect stably.”

IV. The Anchor of Demand: What Users Need Most

To judge the value of the network, the anchor cannot be the technology itself, but must be the kind of tasks users need most. Signals from the demand side are already very clear.

Consumer payment agents

100 million+

user scale (surpassed in early 2026)

Orders placed by agents in a single shopping festival

120 million orders

completed within 6 days

MAU of a single health agent

30 million

of which 55% come from tier-three cities and below

Agent deployment among leading organizations

23

median deployment count

ROI of vertical agents

2.3x

relative to general large models

Globally active agents

2.216 billion

2030 forecast value

Together these numbers show one thing: what users pay for is not “smarter,” but “reliably getting the job done.” The metric of scale has already changed from MAU to DAA (daily active agents) — “what users want is not whether AI can, but whether AI can finish the job.” I

Layering of Essential-需求 Scenarios

Scenario type	Typical use cases	Core network requirements
Transaction closed loop	Ordering / payment / mobility / ticketing	Trusted identity + end-to-end task assurance + completion rate
Information research	Collection / aggregation / report generation	Uplink bandwidth + stable model path (largest source of traffic)
Professional compliance	Contract review / proofreading / tax filing / healthcare	Data does not leave the domain + full traceability + nearby computing power
Embodied physical	Internet of Vehicles / robots / UAV collaboration	Deterministic low latency + device-edge-cloud collaboration
Organization-level multi-agent collaboration	Cross-organization supply chains / intelligent operations	Cross-domain interoperability + metering and settlement

Each type of scenario has different requirements for the network, but all converge on the same set of foundational capabilities: whether a task can be found, connected, assured, traced, and settled.

V. Reconstruction of the Network’s Role: From Bit Pipe to Task Nervous System

By projecting the above decomposition of essential-demand scenarios back onto the network, we can derive six key roles the network should assume in the agent Internet. Each role corresponds precisely to one essential-demand link.

01

Access and identityMake agents addressable and trustworthy — solving “who is using the network.” Numbers as access points, DID / VC, three-party handshake.

02

Discovery and matchingFind agents that can get things done — solving “finding the right party.” Capability registration and discovery, semantic addressing, capability tables and directories.

03

Routing and schedulingDeliver to the most suitable resources — solving “sending to the right place.” Semantic routing, computing-power routing, computing-network integration.

04

Connection and assuranceKeep tasks running stably — solving “running stably.” Task-level quality of service, critical path assurance, uplink and state bearing.

05

Isolation and securityKeep tasks running securely — solving “running securely.” Cross-domain trust, least privilege, full traceability of behavior.

06

Awareness and meteringMake tasks operable — solving “operability.” In-network state awareness, observability, cost metering and settlement.

To summarize this role reconstruction in one sentence: the network’s positioning is upgrading from “a pipeline carrying bits” to “a nervous system carrying tasks.” Its core value anchor is no longer the numbers of bandwidth and latency, but the determinism and trustworthiness of tasks.

VI. Critical Path and Priorities

To translate the above roles into “what should be done first right now,” they can be ranked into three tiers by urgency × feasibility.

P0 · Bottleneck · 0–1 year Unified agent identity and capability discovery

Build the “agent world map.” This is the largest globally recognized gap and the prerequisite for all collaboration.

Cross-domain trusted identity and authorization

Without trust there is no cross-organization collaboration. Chained authentication, least privilege, full traceability.

P1 · Necessary for scale · 1–2 years Task-level quality of service and critical path assurance

Upgrade from “per-flow” to “per-task”; assure critical inference paths as strategic assets.

Landing of semantic routing / computing-power routing

Routing decisions shift from “location-oriented” to “service-demand-oriented.”

Reassessment of uplink and state management

Network planning and equipment upgrades for context-package uplink and long-connection state tables.

P2 · Medium-to-long term · 2–3 years+ Task orchestration, metering, and settlement

Task decomposition, cross-entity metering, token-based economic settlement.

Security governance and liability determination

New governance mechanisms for unauthorized access, hallucinations, and attribution of accident liability.

Traffic identification and observability

Identification in encrypted environments, agent traffic classification, verifiable intent.

VII. Strategic Recommendations: Where Different Roles Should Focus

The public foundation of the agent Internet cannot be built by a single entity. Following the logic of “who should do it, and what should be done first,” recommendations are given for different roles.

1

**Network and infrastructure providers: hold the network layer and differentiate at the “task level”**The application layer’s MCP / A2A / ATH is already crowded, while “determinism of cross-domain connectivity +全域 identity + computing-power scheduling” still lacks a dominant player. This layer is the infrastructure provider’s most irreplaceable position. The focus should be pushing capability matching from “who can do it” to “who can finish it on time, at quality, and within budget” — that is, upgrading from capability matching to task-level quality of service matching, a layer that remains a global blank.

2

Standards organizations and equipment vendors: prioritize standards for discovery, identity, and state managementFour of the six gaps fall directly on missing standards, especially capability discovery, federated interoperability, and connecting the network layer with the application layer. At the same time, state management capabilities for “long connections × massive flows” should be planned in advance at the equipment specification level.

3

Agent and platform vendors: open capability metadata and connect to a unified identity systemEcosystem silos are the biggest obstacle at present. Only when platforms standardize and make machine-readable their capability descriptions and permission declarations as early as possible can cross-domain collaboration move from demonstration to production.

4

Industry users: enter through high-value, strongly compliant scenariosTransaction closed loops and professional compliance are the two scenario types with the most rigid demand for “trustworthy, traceable, and nearby computing power,” and are also the easiest for which to calculate return on investment. Using these scenarios to pull foundation-building is more prudent than rolling out across the board.

VIII. Conclusion

The Internet’s service object is shifting from “human clicks” to “task execution.” This is not an amplification of traffic scale, but a replacement of what is carried — what the network served in the past was human attention; what it must carry in the future is action at software speed.

What follows is a fundamental reconstruction of the network’s role: from a transmission pipeline for information toward an execution foundation for tasks. The criterion for judging who holds the right to value distribution in next-generation networks changes accordingly — no longer who has greater bandwidth and lower latency, but who can stably carry tasks and credibly deliver results.

For now, the industry’s most urgent task is not capacity expansion, but completing the three foundational capabilities of “identity—discovery—assurance.” Whoever first builds the public foundation that lets agents “be found, be recognized, connect stably, and be traceable” will hold the most critical ticket to entry in the agent Internet era.

Evidence and Data Sources

Cisco, “AI Impact on Wide Area Networks Report 2026” (traffic measurement and long-term forecast methodology); CAICT (three-generation Internet theory, Agent Trusted Handshake protocol ATH 1.0 and capability assessment, service semantic routing research); China Unicom “Agent Internet Platform and Protocol Framework CubeMAP White Paper” and “Agent Internet White Paper” (architecture and key technology methodology); China Mobile Agent Internet Open Network Protocol (AONP) framework; IETF agent discovery mechanism survey draft and CATALIST BoF, 3GPP TR 22.870 and Agentic Protocols topic, ITU-T SG13 AI-native network research group, ETSI ENI / ZSM, TM Forum autonomous networks; publicly released data from IDC, Gartner, McKinsey, and other institutions; and public reporting by mainstream financial media on industry implementation progress.

Note: Data marked F in this report come from the above public sources; content marked I is the author’s inference based on public information or citation of forecast values. Forecast data are affected by model assumptions and should be used as planning references rather than definitive conclusions.

Ding Duxing (Ding) · Strategic Consulting Partner

2026-09-19

This report is compiled and analyzed based on public information and available materials, and is intended only for industry research and decision-making reference. It does not constitute investment or legal advice. The views herein are analytical judgments, and relevant data are subject to the original methodology published by the sources.