09 AI Agent Architecture
Lumy is not another chain. Lumy is an AI agent layer.
So this chapter does not cover consensus algorithms or virtual machines. It covers how AI operates beside the user's daily life.
9.1 Why We Defined This as an Agent Layer
The Market Has Changed
Between 2021 and 2023, the value of the blockchain industry was concentrated on "building new chains." Since 2024, the current has shifted. Infrastructure itself is no longer scarce.
What users actually want is not "a faster chain" but "who will handle my assets and daily life on my behalf."
In this new market, value gathers not above infrastructure but closer to the user. We call that place the agent layer.
Three Things the Agent Layer Does
1
Receives user intent in natural language. Users say "organize this week's schedule", they do not call RPCs directly
2
Composes answers by combining multiple external LLMs and tools
3
Executes real actions after explicit confirmation. Payment, settlement, and burning are carried out on the user's behalf after their approval
We Don't Build Models
Large models are already being built by Anthropic, OpenAI, and Google. There is no reason for Lumy to compete in the same arena.
The value Lumy creates lies in tying those models, in one flow, to the user's assets, spending, and daily context.
Built on Standards
Lumy's agent layer is built on an industry-standard interface (Model Context Protocol). Tools written once work identically across multiple models, and new models can join immediately as long as they support the standard.
9.2 The User-Chosen Multi-LLM
The User Chooses the Model on Every Call
For every request, the user selects the model. ChatGPT, Gemini, Claude, and Lumy's own default AI are all available, switchable with a single tap inside the same Lumy AI screen.
If the user does not choose, the request is handled by Lumy's own AI. It is the most cost-efficient option and suits most everyday questions.
When a request needs more capability (image generation, video, music, deep reasoning, long-context analysis), the user picks the right model and sends it there.
The user directly controls both quality and cost. This design treats the user as an adaptive agent.
What the User Sees Is "My LUMY"
Whichever model handles the request, what the user sees is one thing.
The amount of LUMY deducted on this call.
If you don't use it, 0 LUMY; the more calls, the more deducted. Heavier models mean a larger deduction, shown to the user before sending.
What This Design Provides
Unit costs stay low. Aggregated traffic creates leverage in negotiations with multiple vendors.
No single point of failure. Even if one model goes down, another can be picked and the work continues. Modules are loosely coupled, so a shock in one part does not propagate across the whole.
Immediate access to new capabilities. When a new top-tier model that supports the standard launches, it is added to the picker right away.
9.3 Data Permissions and Privacy
Data Tiers Are a Core Architectural Element
The agent handles asset, spending, and daily-context data on the user's behalf. But not all data is of the same tier.
If the system does not enforce this distinction, then a user saying "organize my assets" could result in an external LLM mixing another user's personal information into the answer.
Lumy blocks this risk with four-tier data permissions.
L1
Public
Prices, announcements, Marketplace catalog
L2
Self
The user's own holdings, payments, node status, activity logs
L3
Group
Shared only with explicitly authorized users
L4
Aggregate
Only ecosystem-wide aggregate signals (personal identifiers always masked)
L4 always delivers only aggregates to the user. Individual users' activity is not visible, not even to themselves.
Masking Happens Before the LLM
The user's own identifying information (email, wallet address, account name) and other users' identifiers are masked before reaching the external LLM. What the LLM sees has already been filtered and masked, and answers cannot contain personally identifying information.
Your Own Assets, Your Own Signature
Even when an LLM proposes a real action like a payment or burn, the transaction is always signed by the user themself, in the user's own wallet.
The LLM does not access your keys or seed. "The AI moves my assets however it wants" cannot happen, structurally.
9.4 On-Chain Settlement × Off-Chain Inference
Two Strengths Should Not Sit in the Same Place
The strength of blockchain is trust. The strength of AI inference is performance and flexibility. These two have different requirements.
Response latency
Consensus time (seconds or more)
Immediate (milliseconds)
Mutability
None (permanent record)
Frequent (model/policy swaps)
Cost
Gas per transaction
Processing cost per call
Visibility
Publicly verifiable
Private
What Sits Where
On-chain (smart contracts)
Payment transactions, settlement logic, token burns, node license issuance and exhaustion, holdings lookup
Off-chain (Lumy agent layer)
Natural-language intent parsing, model selection, result summarization and ranking, accumulation of personal activity context
Anything that requires trust is on-chain. Payments, settlements, and burns can be verified by anyone on a block explorer.
What This Separation Guarantees
Because settlement is automated and recorded on-chain, the payment flow is transparently verifiable from end to end. Every path of every payment remains on chain, and every participant can be audited after the fact.
Last updated
Was this helpful?