Exploring the Two-Factory Model for Agent Internet Value Operations

Introduction

Large language models are reshaping software development, but the tension between speed and controllability remains a core engineering challenge. While tools such as Vibe Coding boost productivity, they also amplify quality risks in production environments. This paper introduces Spec‑Driven Development (SDD) as a reusable engineering practice for AI‑assisted development, focusing on the emerging Internet of Agents and the need for operators to shift from a "traffic pipeline" to a "value‑operation hub".

Related Research

The article surveys four strands of prior work: (1) the agent‑economy value chain and the 2A (to Agent) business model, (2) telecom operators’ five‑layer compute model, (3) the Token Factory concept and SCU (Standard Compute Unit) as a standardized accounting unit, and (4) the lack of cross‑organization, full‑network clearing and value‑distribution mechanisms in existing AI‑SaaS solutions.

Two‑Factory Model Architecture

The proposed architecture consists of two parallel factories— Token Factory and Agent Factory —connected by a Value Orchestration Layer . Figure 1 (Token Factory ↔ Agent Factory twin‑star diagram) illustrates the high‑level layout.

Token Factory

Responsible for standardised measurement of AI resource consumption, issuance of Tokens, and asset‑level management. Inputs are compute, network, and power consumption (layers 2‑4 of the telecom compute model); output is a Standard Token anchored to the measured cost.

Agent Factory

Manages the full lifecycle of agents, including DID‑based identity, capability registration, task orchestration, quota control, and marketplaces such as Skill Square and Agent Mall.

Value Orchestration Layer

Translates Agent‑factory task plans into SCU consumption, prices the consumption, and settles value‑added fees using a policy engine. It also enforces governance via a three‑party structure (operator, model vendor, enterprise customer).

SCU Measurement Protocol

The Token Factory defines SCU as the compute required to complete a benchmark inference task on a reference hardware platform. The protocol includes:

Measurement dimensions : input/output token count, context window, task complexity coefficient, latency weight.

Mapping function : normalises heterogeneous model‑hardware consumption to a unified SCU.

Billing events : SCU call‑detail records (task ID, parties, SCU amount, timestamp, model ID), analogous to telecom CDRs.

Operators acquire Token reserves by swapping GPU compute for Tokens (first‑market) and offer locked‑price compute packages to downstream AI service providers (second‑market).

Agent Factory Token‑Pool Mechanism

To avoid “production‑beat‑lag” caused by on‑chain settlement latency, the Agent Factory uses an off‑chain Token Pool (state channel) that enables "execute‑first, settle‑later". The pool supports three credit‑based actions:

Hard block : reject new tasks for low‑credit agents.

Soft downgrade : route tasks to a slower, cheaper model and trigger a recharge alert for medium‑credit agents.

Credit overdraft : allow limited overdraft for high‑credit agents with post‑hoc risk checks.

Four‑Quadrant E‑R‑G‑D Capability Model

The model adds four capability dimensions:

E (Evaluation) : layered offline/online A/B testing and difficult‑case pools, feeding results to the routing module.

R (Routing) : uses SCU forecasts and VPT expectations to allocate Token‑pool quotas, aiming to keep slow‑path usage below 10‑20%.

G (Governance) : DID‑based identity, whitelist, audit logs, VPT anomaly detection, dual‑ledger accounting to prevent fraud.

D (Distribution) : a 2A portal aggregates agents, provides one‑stop onboarding, testing, certification, and billing, and offers value‑optimisation recommendations.

Dynamic Pricing Mechanism

To prevent a single entity from manipulating the SCU‑to‑Token exchange rate, a three‑layer dynamic pricing scheme (Table 1) adjusts the rate based on power cost, GPU depreciation, network overhead, and a profit margin.

Evolution Path for Operators

The paper outlines a three‑stage migration:

Localised entry : Deploy an "Agent‑local Internet" in high‑value verticals (government, energy, industry) with SCU measurement and VPT baseline validation.

Full‑network clearing : Extend SCU standards and VPT evaluation across domains, become a network‑wide AI clearinghouse, and build a Token primary market.

Token‑value operation & ecosystem aggregation : Offer compute‑lock‑in packages, create an Agent credit‑rating system based on VPT, and position the operator as the value‑distribution rule‑maker.

Figure 5 visualises the staged roadmap.

Scenario Validation

A high‑frequency 5G core‑network fault‑diagnosis case demonstrates the model:

Budget: 5 000 Token, value‑event weight: 10 000.

Three agents (core, radio, transport) consume 1 200 SCU, 800 SCU, and 400 SCU respectively (total 2 400 SCU < budget).

VPT = 10 000 / 2 400 ≈ 4.17, exceeding the platform average of 3.2.

Fees: communication = 2 400 × 0.01 CNY = 24 CNY; base value‑add = 3 agents × 2 CNY = 6 CNY; entropy bonus ≈ 1.82 CNY; total ≈ 31.82 CNY.

Revenue split follows contribution percentages (core 50 %, radio 25 %, transport 15 %, platform 10 %).

Abnormal scenarios (VPT = 0, soft‑downgrade, arbitration) illustrate credit‑risk handling, dynamic‑downgrade policies, and dispute‑resolution workflows.

Limitations and Challenges

The authors acknowledge four practical hurdles:

SCU mapping calibration : requires extensive benchmark data; thought‑chain token coefficients lack mature methods.

Token‑pool credit risk : high‑credit overdraft may cause bad debt; VPT volatility can trigger undesired downgrades.

Fast/slow path routing estimation error : VPT forecasts may mis‑estimate new tasks; reinforcement‑learning‑style exploration is proposed.

Four‑token granularity instrumentation : deep framework instrumentation needed; some models (e.g., OpenAI o1) do not expose internal token breakdowns.

Mitigations include tiered calibration, post‑payment credit limits, RL‑based routing, and advocacy for richer model‑vendor APIs.

Conclusion

The "two‑factory" paradigm unifies Token and Agent factories under a telecom‑grade OSS/BSS architecture, introduces a dual‑dimensional SCU‑VPT measurement framework, and provides a concrete three‑phase path from local measurement to network‑wide clearing. By becoming the "value‑measurement standard" and "clearing operator" rather than merely a compute provider, telecom operators can capture the strategic upside of the Internet of Agents.