Inside CarryGo lies a world model—helping organizations see, simulate, act and learn as a system.
At the product level, CarryGo is a world model, built first for business.
At the system level, it is not a single foundation model. It combines a world state, dynamics models, a planner, a durable DAG runtime, Butler–Worker execution and governance.
The distinction matters.
The world model answers: what state is the world in, and how might an action change it?
The planner, DAG and agents answer: which action should be selected, who can execute it, and how can it be completed safely?
Our technical standard for a world model is simple:
It must maintain an updatable world state, estimate action-conditioned state transitions, and continuously calibrate itself against observed reality.
CarryGo connects planning, execution and governance to that model so that simulation can enter real operations.
A business is already a world
A business is made of products, customers, organizations, assets, contracts, channels, rules and relationships.
It is partially observable: nobody has every relevant fact.
It is multi-agent: employees, customers, suppliers, partners, platforms, capital providers and regulators act independently.
It is non-stationary: inventory, cash flow, competition and market sentiment keep changing while decisions are being made.
It is reflexive: once a forecast triggers a price change, campaign, replenishment order or partnership, the object being predicted has changed.
Language models can reason about these dynamics inside a context window. By default, however, they do not maintain a shared, queryable and versioned representation of enterprise reality.
CarryGo is designed to provide that missing layer.
It does more than explain what has already happened. CarryGo brings an enterprise’s resources, market demand, competition, costs and constraints into the same operating world to identify underused capabilities, unmet needs and opportunities worth acting on. See the whole business first; then decide where to begin.
How the business world becomes operational
The system can be understood in five layers:
CarryGo begins with enterprise initialization. Products, orders, inventory, financial metrics, organizational relationships, meeting records, contracts, policies and historical projects enter the same operating model.
They are not merely chunked for retrieval. They become business entities with relationships, time, provenance, confidence and permissions, forming the current World State.
Because a business is partially observable, CarryGo does not claim access to an omniscient ground-truth state. It maintains a belief state Bₜ with evidence, provenance and confidence.
At its core, the modeling problem can be expressed as:
P(Sₜ₊₁ | Bₜ, Aₜ, Eₜ)
Given the current belief Bₜ, a candidate action Aₜ and uncontrollable external events Eₜ, what states are likely to follow? New observations Oₜ₊₁ then update the belief state and calibrate the prediction.
Why CarryGo is a hybrid world model
CarryGo does not rely on one opaque end-to-end model to predict an entire company. It combines four kinds of dynamics:
- deterministic rules for contracts, permissions, inventory constraints, financial limits and approvals;
- probabilistic models for demand, conversion, fulfillment, cost and risk;
- causal and scenario models for estimating the incremental effects of interventions;
- LLMs and multi-agent simulations for proposing hypotheses, generating plans, explaining evidence and modeling stakeholder responses.
Rules enforce hard boundaries. Probabilistic models represent uncertainty. Causal models reduce the risk of mistaking correlation for impact. Agents explore and execute candidate plans in open-ended situations.
The result is a hybrid business world model whose components can be inspected, replaced and calibrated.
From prediction to execution: an illustrative Duowei scenario
Duowei is a Chinese personal-care brand grounded in transparent and intelligent manufacturing, serving everyday care needs across a wide range of scenarios.
Assume conversion in a Duowei livestream falls by 18%. A conventional AI assistant may immediately recommend revising the script or adding a promotion.
CarryGo first asks whether the script is actually the cause.
It reads traffic composition, product clicks, sellable inventory, refunds, fulfillment, pricing and comparable campaigns. It finds three changes: low-fit traffic has reached 41%, the lead product has 1.6 days of inventory, and some demand is being redirected to a lower-margin substitute.
It then simulates two options. The following numbers illustrate the product experience; they are not production performance forecasts.
Option A: maintain media spend, revise the livestream mechanism and replenish inventory immediately.
- Expected incremental contribution profit: RMB 180,000–260,000
- Stockout and fulfillment risk range: 25%–35%
- Key sensitivities: arrival time and effective traffic cost
Option B: reduce low-fit traffic, change the product mix and adjust creator sequencing.
- Expected incremental contribution profit: RMB 120,000–170,000
- Audience-shift risk range: 8%–14%
- Key sensitivities: audience fit and substitute-product margin
After approval, the planner expands the selected option into a project DAG. Content revises the mechanism. Media operations adjust spend. Supply chain validates inventory. The creator’s Butler coordinates scheduling. Finance protects the margin floor. Each node gains an executor, deadline, acceptance criteria and work items.
When outcomes return, CarryGo updates its explanation of the world—not merely a summary. Script quality may lose attribution weight, while audience fit and sellable inventory become more important variables. The next DAG can add a joint audience-and-inventory gate before spend expands.
The DAG describes how CarryGo intends to change the world. The dynamics models estimate how the world may respond. Outcomes record how it actually responded.
That is what evolution means: a more precise graph, a changed structure and a better execution path.
The hard part is not waking up 10,000 agents
Running thousands of model sessions is relatively easy. Allowing agents, people and external systems to change the same business world without corrupting it is much harder.
Every material external action can be understood as an Action Envelope containing at least:
An agent does not modify reality simply because it received a prompt. The action must pass identity, permission, budget, risk and world-version checks.
Durable execution requires leases, idempotency, compensation and snapshots. Governed evolution requires replay, shadow runs, staged release, approval gates, rollback and kill switches.
When we say AI should operate a business, we do not mean unlimited autonomy.
We mean autonomy that can be graded, authorized, observed, interrupted and held accountable.
How we will know whether the model is improving
CarryGo should not measure intelligence by how much its agents say. It should be evaluated against outcomes:
- world-state conflict rate;
- one-step and multi-step state prediction error;
- calibration between predicted risk and observed frequency;
- deviation between expected and realized outcomes;
- constraint violations, duplicate actions and rollbacks;
- first-pass DAG acceptance rate;
- regression failures after model evolution.
If these metrics do not improve, “the model is learning” is only a marketing claim.
How we are building it
CarryGo has established the explicit layer of a business world model: enterprise initialization, business entities and knowledge, context assembly, DAG and project execution, Butler–Worker orchestration, multi-tenant permissions and execution traces.
This layer makes enterprise state traceable, rules expressible and actions executable.
The next stage strengthens learned dynamics: scenario simulation, outcome attribution, forecast calibration and governed DAG evolution.
Beyond that, Butlers representing different organizations can discover opportunities and negotiate under explicit authorization and minimum-disclosure rules. Standardized, low-risk transactions may complete automatically within predefined budget and contract boundaries.
A world model is not a feature completed in a single release. It is a system that becomes more faithful through real operation.
A physical world model helps a robot ask: if I take this step, what will happen around me?
CarryGo helps an organization ask: if we change a product, price, channel, team or partnership, how will the business change next?
LLMs help people work faster.
CarryGo enables an enterprise to understand, predict, act and learn as one coherent world.
