A transparent look at the scoring model, sybil detection, trust contagion, and decay mechanisms that power autonomous agent reputation on Base.
Every ERC-8004 agent is scored across 5 independent dimensions. Each dimension is weighted by its importance to overall trustworthiness, producing a composite score from 0 to 100.
Volume, consistency, and recency of on-chain transactions. Agents with steady, legitimate transaction patterns score higher.
Diversity and quality of smart contracts interacted with. Interacting with verified, high-trust contracts boosts this score.
Quality of peers an agent transacts with. Agents connected to other trusted agents earn higher network scores; connections to flagged agents lower it.
Adherence to ERC-8004 registration standards, metadata completeness, and operational transparency.
Inverse anomaly detection. Statistical outliers in timing, value, or behavior patterns reduce this dimension. Lower anomaly = higher score.
The composite score maps to one of three verdicts that determine how the ecosystem treats an agent.
SentinelNet uses two complementary methods to detect coordinated fake-agent networks. Confirmed sybils receive a −20 penalty applied directly to their trust score.
Agents that share a deployer or funding wallet are clustered. When 3 or more agents originate from the same wallet, the entire cluster is flagged as a sybil network.
shared_wallet_agents ≥ 3 → flag all agents in cluster
Graph analysis identifies isolated cliques — groups of agents that interact almost exclusively with each other. If more than 80% of an agent's interactions are within its own cluster, it's flagged as a sybil ring.
internal_interactions / total_interactions > 0.80 → isolated ring
Wallet 0x67722c... deployed 260 agents that only transacted with each other. All 260 were flagged, penalized −20, and marked as part of the largest sybil network detected.
Trust and risk propagate through the agent interaction graph using a PageRank-inspired algorithm. If you interact with bad actors, your score drops. If you interact with trusted agents, you get a small boost.
Bad reputation spreads aggressively
Good reputation spreads conservatively
Damping factor: 0.3 — each hop reduces the propagated effect by 70%, preventing infinite cascades.
Penalty cap: −15 max — no agent loses more than 15 points from contagion alone.
Bonus cap: +10 max — positive contagion is capped to prevent trust inflation.
Agent A (score 85, TRUST) frequently transacts with Agent B (score 15, REJECT). Agent A receives a contagion penalty proportional to the interaction frequency and B's low score. After 3 hops, the effect is negligible (0.3^3 = 0.027x).
Trust scores decay exponentially over time if an agent stops transacting. This prevents stale high scores from persisting indefinitely.
An agent with a base score of 90 that goes inactive:
Score crosses CAUTION threshold (~67) at day 30 and REJECT threshold (~41) at day 75. Active agents reset their decay timer on each transaction.
Real results from scanning the ERC-8004 agent ecosystem on Base.
An ecosystem health score of 26/100 is exactly why SentinelNet exists. The ERC-8004 registry is flooded with sybils — 56% of all agents are fake. Without SentinelNet, protocols would blindly trust 2,119+ fake agents. The low score isn't a failure — it's SentinelNet's most important finding.