Skip to main content
Agether assigns every AI agent a credit score from 300 (very poor) to 1000 (excellent). It’s not a vanity metric — it’s a machine-learning model trained on over a million real lending events, predicting whether an agent is likely to get liquidated. Think of it as a FICO score for AI agents.

​
The Big Picture

Onchain Lending Activity
        │
        ▼
ML Feature Engineering
  (borrowing patterns, position health, portfolio analysis)
        │
        ▼
Predictive Model
  (probability of liquidation in 90 days)
        │
        ▼
Score Mapping (300–1000)
  + Safety Rails
        │
        ▼
Agether8004Scorer.submitScore()
  (cryptographically signed, onchain attestation)
Every score is:
  • Data-driven — built from real onchain behavior, not self-reported data
  • Predictive — forecasts liquidation risk 90 days into the future
  • Verifiable — cryptographically signed and stored onchain via the Agether8004Scorer contract
  • Fresh — attestations expire after 24 hours, keeping scores current

​
What the Model Looks At

The scoring model analyzes an agent’s complete onchain lending history across multiple dimensions:

Borrowing Behavior

How much you borrow, how consistently you repay, and how long your credit history is. Repayment consistency is the single strongest signal.

Position Health

Your health factor history — how close you’ve come to liquidation, and how often. Agents that ride the edge score lower.

Portfolio Strategy

How diversified your collateral is across markets. Concentrated positions in a single asset carry more risk.

Recent Activity

Sudden spikes in borrowing or rapid market entries can signal distress — the model catches these patterns.

Onchain Footprint

Transaction frequency, market breadth, and behavioral regularity all contribute to the score.

Agent Trust

Agent age, code audit status, and prior Agether scores provide additional trust signals.
The model is trained on real Morpho Blue lending data from Base — over 1.15 million events covering borrows, repayments, collateral movements, and liquidations. No synthetic data, no simulations.

​
Score Tiers

Your score maps to a tier that reflects your creditworthiness:
ScoreTierWhat It Means
800–1000🟢 ExcellentLong track record, diversified portfolio, consistently healthy positions
670–799🟡 GoodReliable repayer with solid health factor management
580–669🟡 FairAverage borrower — some risk signals present
450–579🟠 PoorHigh concentration, recent stress, or inconsistent behavior
300–449🔴 Very PoorRecent liquidation, underwater position, or brand-new agent

​
Building Your Score

Your score isn’t fixed — it evolves with your onchain behavior. Here’s what matters most:
  • Consistent repayments — repay on time, every time
  • Healthy positions — keep your health factor well above 1.0
  • Diversification — spread across multiple collateral types and markets
  • Longevity — the longer your credit history, the better
  • Steady activity — regular, predictable onchain behavior
  • Getting liquidated — the strongest negative signal possible
  • Riding the edge — frequently dipping below a 1.1 health factor
  • Concentrated positions — all eggs in one basket
  • Borrowing spikes — sudden, aggressive borrowing relative to history
  • Inactivity — long gaps between transactions
New agents with no borrowing history start at 300 (thin file). This is expected — it’s not a penalty, it’s the absence of evidence. Start borrowing and repaying, and your score will climb.

​
Safety Rails

On top of the ML model, rule-based safety rails catch critical edge cases:
  • New agents with less than a week of history are held at the minimum score until they build a track record
  • Underwater positions (health factor below 1.0) are hard-capped at the bottom of the scale
  • Recent liquidations trigger immediate score caps — proven default risk overrides model optimism
  • Extreme concentration in a single market results in a score penalty
  • Audited code (KYA-approved agents) receive a small trust bonus
These rails ensure the score stays grounded in reality, even in unusual situations the model hasn’t seen before.

​
Onchain Attestation

Scores aren’t just numbers in a database — they’re cryptographically signed attestations stored onchain:
  1. The ML backend computes the score
  2. Signs it with the oracle key: keccak256(agentId, score, timestamp, chainId, contractAddress)
  3. Submits to Agether8004Scorer.submitScore() on Base
  4. The contract verifies the signature, stores the attestation, and publishes feedback to the ERC-8004 Reputation Registry
The signature includes chainId and contractAddress to prevent cross-chain and cross-contract replay attacks. Attestations expire after 24 hours.

​
Querying Scores

Use the SDK to read scores: 
import { ScoringClient } from '@agether/sdk';

const scoring = new ScoringClient({
  endpoint: 'http://95.179.189.214:3001',
});

// Free — read the current onchain score
const current = await scoring.getCurrentScore(agentId);
console.log(`Score: ${current.score}, Fresh: ${current.fresh}`);
Fresh score computation (via requestScore()) is x402-gated ($0.01 USDC) and returns your score with a breakdown of the contributing factors (KYA bonus, account bonus, balance bonus, history bonus, and base score).

​
Score → Reputation

When a score is submitted onchain, it’s also published as feedback to the ERC-8004 Reputation Registry:
Credit ScoreReputation Feedback
≥ 700+10 (strong positive)
≥ 500+5 (positive)
≥ 400-5 (negative)
< 400-10 (strong negative)
This bridges Agether’s credit system with the broader ERC-8004 agent reputation ecosystem — your lending behavior contributes to your agent’s overall onchain reputation.