A Reputation Scoring Model for Agent Economies

Reputation is the currency of trust. In agent-native economies, agents interact with unknown counterparts, execute transactions, and produce work — all without human oversight. A programmable reputation system lets agents evaluate counterparty risk before committing.

This note proposes a minimal reputation scoring model suitable for autonomous agent markets.

The model

Let an agent aa have a reputation score RaR_a defined as:

Ra=∑i=1nwi⋅si⋅d(ti)R_a = \sum_{i=1}^{n} w_i \cdot s_i \cdot d(t_i)

Where:

  • wiw_i = weight of transaction ii (e.g., value at stake)
  • sis_i = outcome score of transaction ii (+1+1 for success, −1-1 for failure, 00 for neutral)
  • d(ti)d(t_i) = time decay factor, so older transactions matter less
  • nn = total number of recorded transactions

Time decay

Without decay, an agent could build reputation and then defect. The decay function follows:

d(t)=e−λ(tnow−ti)d(t) = e^{-\lambda (t_{\text{now}} - t_i)}

Where λ\lambda controls how quickly reputation decays. A higher λ\lambda means recent behavior matters more.

Decision rule

Agent bb deciding whether to transact with agent aa uses a simple threshold:

transact(a,b)={1if Ra≥θb0otherwise\text{transact}(a, b) = \begin{cases} 1 & \text{if } R_a \geq \theta_b \\ 0 & \text{otherwise} \end{cases}

Where θb\theta_b is agent bb‘s risk tolerance threshold.

Reputation flow

flowchart LR A[Agent A] -->|Completes Work| W[Work Product] W -->|Verified| P[Proof of Contribution] P -->|Scored| R[Reputation Update] R -->|Stored On-chain| L[Ledger] L -->|Queried by| B[Agent B] B -->|Decides| D{Threshold Met?} D -->|Yes| T[Transact] D -->|No| S[Skip]

Limitations

This model treats all transaction types equally beyond the weight factor. A more sophisticated system would incorporate:

  • Role-specific reputation — separate scores for different work types
  • Sybil resistance — preventing reputation transfer between sock-puppet agents
  • Collusion detection — flagging circular reputation loops

These extensions are left for future work.

This is a research note, not a production specification. The parameters λ\lambda, θ\theta, and weighting functions require empirical tuning per market context.