Agent Trust Score¶

The ChainAware Agent Trust Score is a 0-1000 score that measures how safe it is to interact with any ERC-8004 registered AI agent.

Unlike voting-based reputation systems - where agents can upvote each other to manufacture trust - the Agent Trust Score is derived entirely from on-chain behavioral history. It cannot be earned in hours. It cannot be faked with a cluster of fresh wallets. It reflects the real-world track record of the human or entity controlling the agent.

As agentic commerce scales - with AI agents autonomously completing purchases on behalf of consumers across ChatGPT, Google Gemini, and Shopify - the question of which agents can be trusted to transact is no longer theoretical. ChainAware answers it with on-chain evidence, not peer endorsements.

Score range: 0-1000
Five tiers: Untrusted → Elevated Risk → Provisional → Trusted → Sovereign
Chains supported: Ethereum, BSC, Base, Avalanche, Mantle

Agentic Commerce and the Trust Gap¶

Agentic commerce - where AI agents autonomously discover, compare, and purchase products on behalf of consumers - is projected to account for 25% of global online retail spend by 2030. Major platforms including ChatGPT, Microsoft Copilot, and Google Gemini are already executing real transactions. McKinsey estimates the model could redirect $3-5 trillion in global retail spend by the end of the decade.

This creates a fraud surface that traditional detection cannot handle. AI agents exhibit behavior patterns - rapid sequential orders, cross-category purchases, unusual velocity - that legacy fraud systems flag as suspicious, generating false positives that block legitimate agents while missing sophisticated bad actors entirely. 78% of financial institutions expect fraud to spike specifically from AI shopping agents operating in agentic commerce channels.

The missing layer is agent identity. Know Your Agent (KYA) protocols are emerging as the commerce equivalent of KYC - but most implementations stop at cryptographic identity verification. They confirm the agent exists. They do not tell you whether the human controlling it has a history of rug pulls, mixer exposure, or Sybil farming.

ChainAware's Agent Trust Score is the KYA layer built on behavioral intelligence - tracing the agent back to its owner wallet, its payment address, and the funding source behind both, then scoring that history against 20M+ wallet personas and one year of on-chain fraud data.

Why Voting-Based Agent Trust Fails¶

ERC-8004 includes a built-in Reputation Registry - a standard interface for agents to receive peer feedback. The design is intentionally open: anyone can leave a review, and scoring is left to third parties.

This creates a well-known attack vector. An operator deploys 50 agent wallets. Each one reviews every other. All 50 accumulate high reputation scores within hours. The cost: near zero. The result: manufactured trust indistinguishable from genuine reputation.

In an agentic commerce context, this is not a theoretical risk. An agent with a manufactured reputation score can be deployed as a buyer, seller, or intermediary in autonomous transaction flows - and no voting-based system will catch it before funds move.

ChainAware does not read the ERC-8004 Reputation Registry to compute the Agent Trust Score. We look behind the agent at the wallet that controls it - and the wallet that funded that controller - and score what those wallets have actually done on-chain over their entire history.

The Three Pillars¶

Every Agent Trust Score is built on the same three pillars as the ChainAware Wallet Reputation Score:

Experience
How long and how actively has the owner wallet operated on-chain? A wallet with years of diverse transaction history across protocols carries more weight than one created last week.

Risk Capability
What is the behavioral profile of the owner wallet? Does it interact with a range of DeFi protocols, or does it show narrow, mechanical patterns consistent with automated fraud?

Fraud Probability
ChainAware's predictive AI model scores the likelihood that the owner wallet will engage in fraudulent behavior, trained on 20M+ wallet personas across ETH, BSC, Base, and beyond. This is not a blacklist check - it is a forward-looking prediction.

These three pillars feed the same formula used by the Wallet Reputation Score:

ReputationScore = (1000/110) × (experience + 1) × (risk_capability + 1) × (1 − fraud_probability)

Maximum: 1,000. The Agent Trust Score applies this formula to both the owner wallet and the agent payment wallet, then combines them using trust delegation logic described below.

What We Score¶

The Owner Wallet (Agent Creator)¶

The wallet that registered the agent on the ERC-8004 Identity Registry. This is the controlling entity - the human or organization behind the agent. We score its full behavioral history using ChainAware's fraud prediction model.

The Agent Wallet¶

The payment address where the agent receives funds. This is often a fresh address with no history. The Agent Trust Score does not penalise a fresh agent wallet if the owner wallet has strong trust - instead, the owner's reputation is partially inherited by the agent wallet through our trust delegation mechanism.

The Feeder Address¶

The wallet that funded the owner. This is ChainAware's most distinctive signal - no other agent trust platform traces agent ownership back to its funding source. If the owner wallet was funded by a known fraud actor, mixer, or prior rug pull operator, that signal propagates to the agent score regardless of how clean the owner wallet appears on its own.

Feeder analysis is available for approximately 38% of indexed agents. When the feeder is a verified CEX withdrawal address (Binance, Coinbase, Kraken, and others), this is treated as a strong legitimacy signal - CEX withdrawals imply the owner passed KYC somewhere upstream. When the feeder is unknown or obfuscated, the score reflects this uncertainty with a penalty.

Prior Token and Pair History¶

ChainAware maintains a database of token creation and liquidity pair history across one year of on-chain activity. We cross-reference the owner and feeder addresses against this database to identify:

• Honeypot tokens previously created by the agent owner or feeder
• Rug pull pairs - liquidity pools where the creator removed funds in patterns consistent with exit fraud

A single confirmed rug pull or honeypot in an agent owner's history results in a hard score cap regardless of all other signals. This is not adjustable. An agent controlled by a serial rug puller cannot score above the minimum tier.

This cross-product signal is unique to ChainAware: an operator who rugged a token in Q4 2025 and registered 47 agents in Q1 2026 is caught by both the Rug Pull Detector and the Agent Trust Score. No other platform connects these datasets.

Trust Delegation¶

Agent wallets are frequently fresh addresses created specifically for the agent. A naive scoring approach would penalise every new agent wallet - producing low scores for legitimate agents and making the score less useful in agentic commerce contexts where new agents are deployed continuously.

ChainAware solves this with trust delegation: the owner wallet's reputation sets a floor for the agent wallet's effective score. A strong owner partially transfers credibility to a fresh agent wallet. A fraud-flagged owner cannot transfer any credibility - the delegation factor collapses to near zero.

This means:

• A reputable developer launching their first agent scores appropriately high, even with a fresh payment wallet
• A fraud-flagged wallet launching an agent scores near zero regardless of how clean the agent wallet appears

Farm Detection¶

ERC-8004 does not prevent one owner from registering hundreds of agents. In agentic commerce, agent farms are a specific attack vector: operators flood marketplaces with low-trust agents that cross-endorse each other, creating the appearance of ecosystem depth while controlling the supply for manipulation or fraud.

ChainAware tracks agent fleet size per owner across all indexed chains. Owners controlling large numbers of agents - particularly those registered in the same block - receive a farm detection signal that suppresses the score for every agent in the fleet.

This catches the most common Sybil pattern in the ERC-8004 ecosystem: operators creating clusters of agents designed to cross-review each other, interoperate artificially, or flood agentic commerce marketplaces with low-quality supply.

EIP-7702 Delegation¶

Some agent owners use EIP-7702 to delegate control of their wallet to a secondary address. ChainAware detects this and scores both the registered owner and the delegate. The Agent Trust Score takes the lower of the two - if the delegate is a fraud-flagged address, the agent score reflects that regardless of the registered owner's history.

Approximately 5% of indexed agents have EIP-7702 delegated ownership. These are flagged explicitly in the API response.

When to Call the Agent Trust Score¶

ChainAware can be called at any point in the ERC-8004 agent lifecycle:

• At registration - screen new agents before they're listed in the Identity Registry
• At discovery - enrich reputation data with behavioural intelligence before hiring
• Pre-transaction - screen agent wallets in real time before each execution
• Periodic review - re-screen listed agents monthly as their wallet history evolves

A typical agent-to-agent hiring flow:

Agent A wants to hire Agent B for a task
  → Looks up Agent B in the Identity Registry (ERC-8004)
  → Checks Agent B's Reputation Registry score (ERC-8004)
  → Calls ChainAware Agent Trust Score on Agent B's wallet + feeder wallet
  → Gets fraud score, AML status, behavioural verdict
  → Makes an informed hire/no-hire decision

Score Tiers¶

What the API Returns¶

The Agent Trust Score API returns the score, tier, and a set of flags. Flags describe the category of signal detected - they do not expose the thresholds or weights behind them.

What We Do Not Publish¶

ChainAware publishes the categories of signals used to compute the Agent Trust Score. We do not publish the exact weights, thresholds, or model coefficients. This is deliberate.

Publishing exact thresholds would allow bad actors to calibrate their behavior to stay just below each cap. Publishing the fraud model coefficients would allow adversarial wallet warming to target specific score improvements.

Our fraud prediction model retrains continuously on new on-chain patterns. Even partial knowledge of current weights becomes stale quickly.

What cannot be gamed regardless of formula knowledge: immutable on-chain history. A wallet that created a honeypot token in November 2025 cannot remove that event from the blockchain. A feeder address that funded three rug pull operators cannot change its transaction graph. In agentic commerce, where agents transact autonomously at scale, this immutability is the foundation of meaningful trust - not reputation scores that can be manufactured overnight.

How It Differs From Competitors¶

Six platforms now operate in the agent trust space: ERC-8004 native, RNWY, SkyeProfile, AXIS T-Score, DJD Agent Score, and ChainAware. They answer different questions - peer review quality, wallet activity, task performance, or behavioral fraud history - and are largely complementary rather than directly competitive.

ChainAware is the only platform that traces agent ownership to the feeder address (the wallet that funded the owner), cross-references both against a rug pull and honeypot token database, and runs a predictive fraud model before any interaction has taken place.

Full platform comparison →

API Access¶

GET /erc8004/agent/{chain_id}/{agent_id}/trust-score

Returns:

{
  "chain_id": 1,
  "agent_id": 4471,
  "agent_trust_score": 87,
  "tier": "Untrusted",
  "flags": [
    "CREATOR_RUG_HISTORY",
    "FARM_DETECTED",
    "FEEDER_UNKNOWN"
  ],
  "scored_at": "2026-06-25T09:14:00Z"
}

Free tier available - no signup required for public agents.

Related Products¶

Rug Pull Detector - The same creator chain traversal logic that powers rug pull detection feeds directly into the Agent Trust Score criminal record check. An operator who rugged a token in Q4 2025 and registered agents in Q1 2026 is caught by both products.

Wallet Reputation Score - The Agent Trust Score uses the same 0-1000 formula as the Wallet Reputation Score, allowing direct comparison between a wallet's standalone reputation and the trust level of agents it controls.

Fraud Detection API - The underlying predictive fraud model used to score owner and feeder wallets. The same engine powering Know Your Agent intelligence for agentic commerce platforms.

chainaware-agent-screener - A ready-made Claude Code subagent that handles the full dual-wallet screening in a single prompt. Screens both the agent wallet and its feeder wallet, applies the trust formula, and returns a normalized verdict - no API integration required for teams already using Claude Code.