ERC-8004 tells you what an agent declares about itself; it says nothing about how the agent actually behaves Why verified identity alone does not establish agent trust explains the distinction in more depth. — and that gap is where fraud lives. ChainAware is named in the CB Insights AI Fraud Prevention Market Map for closing this behavioral intelligence gap.
When teams evaluate AI agents at scale, they need a number they can act on. Among the inputs, the funding chain is often the most revealing, and our analysis of feeder wallet risk in autonomous agents shows why it deserves close attention. A score is only as useful as the methodology behind it, and a methodology is only defensible when every input is named, weighted and updated on a known schedule. This article sets out ChainAware’s full approach to calculating the Agent Trust Score for ERC-8004 agents To see the signals in practice, the weekly ERC-8004 fraud pattern tracker reports how farm and rug-pull linked agents appear across the registry.: what the registry supplies, what ChainAware adds, how the five behavioral signals combine into a 0-1000 composite and what the score deliberately leaves out.
If you are building a checkpoint into an agentic pipeline, integrating trust signals at checkout For a broader overview of how the score fits into agentic commerce, see our AI agent trust score product overview. Teams applying these scores at payment time may also want to review how The predictive_fraud tool delivers real-time AI fraud probability and AML forensics for any wallet, which complements agent-level checks at payment time. tokens, mandates and behavioural trust work together at checkout. or doing technical due diligence on an agent trust platform, this is the reference document. For the complete technical specification — including the scoring formula, hard caps, trust delegation mechanics and API response schema — see the Agent Trust Score learn page.
What ERC-8004 Registers — and What It Does Not
ERC-8004 is an on-chain agent registry standard. When an agent registers, the record contains four fields The agents_trust_score_single tool returns a full trust profile for any single ERC-8004 agent, including its 0-1000 score and trust flags.: the agent’s wallet address, the creator address that deployed it, the owner wallet that controls it and the mandate scope, which is a declared description of what the agent is authorised to do.
That is the full extent of what the registry provides. The predictive_behaviour tool profiles observed wallet activity to show how an agent actually behaves, filling the gap the registry leaves open. As the ERC-8004 trust signals article explains in detail, registration is a declaration, not a proof. Any operator can register an agent with a clean-sounding mandate and a fresh wallet. The registry has no mechanism to verify that declared scope matches actual behaviour, that the owner wallet is not linked to past misconduct or that the agent is not one of thousands deployed from the same Sybil farm.
This is not a flaw in ERC-8004. Identity registration and behavioural trust are different problems. The registry solves the first. ChainAware’s Agent Trust Score solves the second. As the KYA vs agent identity article argues, knowing who an agent is does not tell you whether to trust it.
The Five Behavioral Trust Signals
ChainAware indexes every ERC-8004 agent and constructs five independent signals. Each signal is scored separately, then combined into the composite. The signals are: funding chain, fleet detection, mandate compliance, transaction velocity and creator reputation.
1. Funding Chain
The feeder wallet — the address that funds an agent’s operating wallet — is the single highest-signal input in the model. ChainAware traces the full funding graph upstream from the agent’s wallet, checking every node against OFAC sanctions lists, confirmed rug pull operator addresses, mixer outputs and flagged infrastructure wallets.
Of the 1,648,406 agents currently indexed, 1,107,688 (67.2%) have feeder addresses that are entirely unknown — no traceable CEX withdrawal, no attributed on-chain history. A further 930 are funded by wallets tied directly to confirmed rug pull operators, and 116,083 (7.04%) have feeders classified as fraud by the behavioral model. Those agents receive a near-zero funding chain sub-score regardless of anything else in their record. The funding-chain risk article walks through how that graph traversal works and why feeder address contamination is the fastest route to a failed agent trust check.
Funding chain is the only signal updated in real time. When a feeder wallet is flagged, every agent it touches re-scores within minutes.
2. Fleet Detection
Legitimate operators deploy agents in small, purposeful sets. Sybil farms deploy them in large, structurally similar clusters. Fleet detection applies network graph clustering to creator addresses, shared infrastructure patterns, deployment timing and wallet funding topology to identify agents that belong to coordinated farms.
At the most recent measurement, 91,868 agents (5.57%) show FARM_DETECTED flags, meaning a single operator registered 50 or more agents, or registered five or more in the same block. Membership in a confirmed farm does not automatically produce an Untrusted score, but it applies a significant negative weight that an otherwise clean record must overcome. Fleet detection updates hourly, because farm operators respond quickly to detection and re-register under new addresses.
3. Mandate Compliance
An agent’s ERC-8004 record declares a mandate scope: what the agent is authorised to do. ChainAware compares that declaration against on-chain activity. An agent registered as a narrow treasury rebalancer that initiates NFT purchases and cross-chain bridge calls is operating outside its declared scope. Each out-of-scope action type is logged and reduces the mandate compliance sub-score.
Mandate compliance matters most to enterprise integrators and acquirer tech teams, because it is the signal most directly tied to whether an agent is doing what its operator claimed it would do. It is also the signal most useful for autonomous compliance screening, where declared scope is a contractual commitment.
4. Transaction Velocity
Every use case has a natural transaction rate. A personal finance agent executing a few daily rebalancing trades has a different expected velocity from a high-frequency arbitrage agent. ChainAware builds a velocity baseline for each declared mandate category and flags agents whose transaction rate is inconsistent with their stated purpose.
Velocity anomalies surface two risk patterns. The first is agents operating far above their declared use case, suggesting either undisclosed activity or a compromised agent being used as infrastructure. The second is agents with near-zero activity despite active registration, which correlates with Sybil accounts maintained for future activation. The agent fraud patterns article documents both in detail using weekly tracker data.
5. Creator Reputation
A creator address that has previously deployed agents rated Untrusted carries that history forward. Creator reputation aggregates the trust distribution of all agents deployed from a given creator address and applies a prior to new deployments from the same source. For more detail, see fraud pattern data across 1.6M agents.
This signal is the most consequential for new agents. A freshly registered agent has no behavioral history of its own, so the creator record is the dominant input at launch. It is also why median trust scores for new agents are low: at Week 41 2026, the median trust score of newly registered agents was 101, partly because many were deployed from creators with poor prior records.
Creator reputation updates daily, reflecting the full rolling history of each creator address.
Compositing Into a 0-1000 Score
The five signals combine into a weighted composite. Funding chain carries the heaviest weight because contaminated funding is the clearest indicator of operator-level misconduct. Fleet detection and creator reputation are the next tier, because they reflect systemic patterns that individual agent behaviour cannot override. Mandate compliance and transaction velocity carry lower base weights but can apply large penalties when anomalies are severe.
The composite runs from 0 to 1000 and maps to five tiers.
Sovereign: 800 to 1000. Clean funding chain, no fleet association, strong creator history, consistent mandate compliance and velocity within expected range. These agents receive minimal friction at integration checkpoints.
Trusted: 600 to 799. One or more signals show minor variance but nothing that indicates active misconduct. Standard monitoring is appropriate.
Provisional: 400 to 599. Multiple signals are ambiguous or the agent is too new to have resolved uncertainty. Additional review or capability restriction is recommended before granting full access.
Elevated Risk: 200 to 399. At least one signal shows a clear negative indicator. Agents in this tier account for the dominant share of the indexed population: 73.5% of all scored agents sit in this tier alone. Combined with the Untrusted tier, 82.5% of all indexed agents score Elevated Risk or Untrusted.
Untrusted: 0 to 199. Direct funding chain contamination, confirmed farm membership or severe mandate violation. Block or escalate. For a comparison of how this tier structure maps against other approaches, see the agent trust platforms comparison.
Live Tier Distribution
As of 2026-11-09, the tier distribution across 1,648,406 indexed ERC-8004 agents is:
| Tier | Score Range | Agent Count | Share |
|---|---|---|---|
| Sovereign | 800–1000 | 130,648 | 7.9% |
| Trusted | 600–799 | 73,110 | 4.4% |
| Provisional | 400–599 | 82,541 | 5.0% |
| Elevated Risk | 200–399 | 1,212,017 | 73.5% |
| Untrusted | 0–199 | 148,096 | 9.0% |
The dominant tier by far is Elevated Risk, driven heavily by the 1,107,688 agents (67.2%) whose feeder address is unknown. The top platform by agent volume is centralcity.ai with 1,051,344 agents (average trust score 258), followed by evoevo.ai with 143,039 (avg 427) and termix.ai with 88,917 (avg 514). These platform-level averages reveal how operator quality, not agent-level behavior, drives the bulk of the score distribution. For a full breakdown of how trust scoring works across Base chain’s 1,175,319 agents — the largest single-chain registry — see the dedicated Base chain analysis.
Update Cadences
Different signals change at different speeds, so each updates on its own schedule.
Funding chain: real-time. Sanctions lists and confirmed operator flags propagate immediately to every agent in the affected graph.
Fleet detection: hourly. Farm re-registration is fast, so detection must be fast.
Mandate compliance and transaction velocity: daily. These signals require a window of observed activity to be meaningful.
Creator reputation: daily. Aggregated across all deployments from a creator address.
What the Score Does Not Capture
Honest methodology documentation names its limits. The Agent Trust Score does not assess subjective intent. It measures observable signals, not whether an operator meant to cause harm. Two agents can produce identical scores through very different paths.
The score also does not capture off-chain activity. If an agent coordinates with counterparties through private channels or executes logic entirely within a system that does not write to a monitored chain, that activity is invisible to the model. This is a structural limit of on-chain behavioral analysis, not a gap specific to this methodology.
Finally, the score is a risk indicator, not a legal determination. It does not replace compliance review for regulated use cases; it prioritises which agents require that review.
Where the Trust Score Sits in the Agentic Economy Stack
A trust score that only lives inside one platform is not a trust score — it is a private opinion. The ERC-8004 standard was designed with that distinction at its center. The score is not a proprietary risk flag computed in a closed system and consumed only by the entity that produced it. It is an on-chain, standardized signal that any protocol, payment rail, enterprise gateway, or API marketplace can read, weight, and act on. That interoperability property is what makes the ERC-8004 trust score the foundational Know Your Agent behavioral layer for agentic economy infrastructure, rather than just another fraud-detection output.
To understand why that distinction matters, consider what the agentic economy actually requires to function. Autonomous agents transact across chains, across protocols, and across organizational boundaries — often within the same workflow. A DeFi protocol, an enterprise procurement system, and a third-party API marketplace may all interact with the same agent address within minutes. If each of those surfaces maintains its own siloed behavioral model with no shared signal, the agent can be low-trust on one platform and fully credentialed on another, and the arbitrage between those two states is exactly where systemic risk accumulates. A shared, on-chain trust layer closes that gap by giving every participant access to the same behavioral record, without requiring them to share raw transaction data or coordinate off-chain.
How Trust Scoring Compares to Existing Risk Frameworks
It is worth being precise about what makes agent trust scoring structurally different from the risk frameworks that preceded it, because the differences are not cosmetic — they reflect a genuinely new threat and governance surface.
Credit scores are static identity instruments. They aggregate historical repayment behavior into a number that describes a human financial identity at a point in time. They are designed for infrequent, high-stakes decisions — a mortgage approval, a credit card limit — and they assume a stable, legally accountable subject. Agents are none of those things. They can be deployed and deprecated in hours, they carry no legal identity, and their behavior can change between one block and the next.
Fraud scores operate at the transaction level. They evaluate a single event — a payment, a login, a transfer — against a model of normal behavior and return a binary or probabilistic risk flag. They are fast and effective for point-in-time anomaly detection, but they have no memory across sessions, no awareness of the agent’s fleet context, and no model of how funding sources connect to behavioral patterns. Each transaction is evaluated in isolation.
The ERC-8004 trust score is continuous, compositional, and structurally aware. It incorporates behavioral trajectory across multiple time windows, fleet-level signals that connect individual agent addresses to their operator clusters, funding chain provenance that traces capital back through the deployment graph, and cross-protocol consistency that detects when an agent behaves differently depending on which platform it believes is watching. No existing credit or fraud framework operates across all four of those dimensions simultaneously, and no existing framework is designed to be consumed as a shared, interoperable on-chain primitive. That combination is what makes trust scoring a new category, not an incremental improvement on existing tools.
How the Score Is Used Across Agentic Commerce Contexts
In practice, the ERC-8004 score surfaces at several distinct points in the agentic commerce stack, and the access control logic differs meaningfully across each of them.
DeFi protocols use the score as a gating condition at the liquidity access layer. An agent requesting access to a lending pool, a yield strategy, or a cross-chain bridge can be evaluated against a minimum trust threshold before the transaction is processed. Protocols can configure that threshold dynamically — tightening it during periods of elevated fleet-level anomaly activity and relaxing it for agents with long, consistent behavioral histories. This is not blacklisting; it is continuous, graduated access control that reflects actual behavioral evidence rather than a static whitelist.
Payment authorization systems use the score at the moment of transaction initiation. For high-value or cross-border agent-initiated payments, the trust score functions as a pre-authorization signal that determines whether the transaction routes through standard rails or is flagged for additional review. This is structurally similar to how card networks use velocity and behavioral signals for human cardholders, but applied to an agent whose behavioral profile is richer and more continuous than any human payment history.
Enterprise agent onboarding uses the score as a credential substitute. When an organization is evaluating whether to grant an external agent access to internal APIs, procurement workflows, or financial systems, the ERC-8004 score provides a standardized behavioral attestation that the enterprise does not have to generate itself. Rather than running a bespoke evaluation of an unknown agent address, the enterprise reads the on-chain trust record and makes an admission decision against its own configured thresholds.
API marketplaces use the score for tiered access control. Agents with higher trust scores gain access to higher-rate-limit tiers, more sensitive data endpoints, or reduced friction authentication flows. Agents with thin or anomalous trust histories are routed to restricted tiers or required to complete additional behavioral validation before gaining full access. The score becomes the passport that determines what an agent can reach within the API economy.
The Network Effect and the Self-Regulating Agentic Economy
There is a compounding dynamic embedded in the architecture of a shared trust layer that is not present in any proprietary risk system. When a single
API Access via Prediction MCP
The Agent Trust Score is available through two Prediction MCP tools. agents_trust_score_single returns the full score and signal breakdown for one agent address. agents_trust_score_list accepts a batch of addresses and returns scores for all of them in a single call, which is the right tool for fleet-level screening at intake or for periodic re-scoring of an existing approved list.
Both tools return the composite score, the tier label, each sub-signal score and the timestamp of the last update for each signal. The response structure is documented in the Prediction MCP setup guide.
The free tier supports 1,000 queries per month, which is enough to instrument a development environment or validate the methodology against a known agent population before committing to a production integration.
QUERY THE SCORE
Score any ERC-8004 agent in seconds with Prediction MCP.
Use agents_trust_score_single or agents_trust_score_list to retrieve the full behavioral composite, tier label and per-signal breakdown for any agent or fleet. Free tier includes 1,000 queries per month.
ERC-8004 Methodology
Query the live Agent Trust Score for any registered agent