Agent Architecture

Full Agent Layer Diagram

graph TD
    subgraph EXTERNAL["External Data"]
        CL["Chainlink Coffee/USD"]
        GFW["Global Forest Watch API"]
        WX["OpenWeatherMap\nUganda drought index"]
        AT["Africa's Talking\nUSSD gateway"]
    end

    subgraph AGENTS["ERC-8004 Agent Layer (Mantle)"]
        RM["LendingVault\nRisk Monitor"]
        USSD["USSD\nSession Agent"]
        DDS["DDS Anomaly\nDetector"]
        CI["Cooperative\nIntelligence Agent"]
    end

    subgraph CONTRACTS["Smart Contracts (Mantle chainId 5000)"]
        FR["FarmerRegistry.sol"]
        BT["BatchToken.sol"]
        TL["TraceLog.sol"]
        LV["LendingVault.sol"]
        CS["CreditScore.sol"]
        PO["PurchaseOrder.sol"]
        PF["ProtocolFee.sol"]
    end

    subgraph AUDIT["Audit Layer"]
        HCS["Hedera HCS\nAppend-only log"]
        ERC["ERC-8004 Registries\nIdentity / Reputation / Validation"]
    end

    subgraph OUTPUT["Outputs"]
        DASH["Cooperative Dashboard"]
        FB["Fonbnk\nMTN MoMo payout"]
        TRACES["EU TRACES\nDDS filing"]
    end

    CL --> RM
    WX --> RM
    RM --> LV
    RM --> HCS
    RM --> ERC

    AT --> USSD
    USSD --> FR
    USSD --> BT
    USSD --> FB
    USSD --> HCS

    TL --> DDS
    GFW --> DDS
    DDS --> TRACES
    DDS --> HCS

    CS --> CI
    LV --> CI
    BT --> CI
    TL --> CI
    HCS --> CI
    CL --> CI
    CI --> DASH
    CI --> ERC

Design Principles

Agents read, contracts write. No agent can modify contract state directly. Agents read from contracts via public view functions and external APIs. They submit transactions through the same access-controlled API layer that human operators use. The on-chain security model is unchanged.

Every decision is recorded before execution. An agent recommendation is written to Hedera HCS and the ERC-8004 Validation Registry before any action is taken. If the action fails, the recommendation record still exists. This creates an audit trail of intent, not just outcome.

Agents have identities, not just addresses. Under ERC-8004, each AsiliChain agent has an NFT-based identity on Mantle mainnet. Its decisions accumulate against a reputation score. MFIs evaluating the protocol can query an agent’s historical recommendation accuracy before trusting it with their capital.

No autonomous high-stakes execution in Phase 1. Liquidation, forbearance approval, and emergency pause all require human multisig in Phase 1 and Phase 2. Agents provide the inputs to those decisions — they do not make them unilaterally. Autonomous execution of high-stakes actions is a Phase 3 capability, gated on an established on-chain track record.

ERC-8004 Identity Registry (Deployed)

AsiliChain deployed its own ERC-8004-compliant IdentityRegistry on Mantle Sepolia (chain 5003):

Proxy:       0x62a6b58f8c3625F0c5f46D6C86A65595AA769C89
Implementation: 0x2E9D3EDf3406c4bd1a8f2CAE814c35D2294Fa02e
Owner:       0xB70f03dE20c9D4c90246c830F81D44f377A652C0

Registered agents:

Agent	Agent ID	Token Owner
risk-monitor	0	deployer
anomaly-detector	1	deployer

Reputation and Validation Registries are not yet deployed. AsiliChain agents currently register on the Identity Registry only — reputation will be earned on-chain as agents execute cycles.

Agent Technology Stack

Layer	Technology
Agent runtime	Python 3.12, CrewAI framework
LLM for reasoning	Configurable (MiniMax M1, Claude, GPT-4)
On-chain reads	viem.js public client, Mantle mainnet RPC
On-chain writes	API wallet relayer (same gas sponsorship as human ops)
ERC-8004 registration	ethers.js, Mantle mainnet
HCS writes	@hashgraph/sdk
Scheduling	Node.js cron (15-min / 30-min cycles)
External APIs	Chainlink, GFW, OpenWeatherMap, Africa’s Talking
Agent endpoints	Next.js API routes at agents.asilichain.xyz