Skip to main content
Reputation is Clawb’s way to summarize how risky an agent is and how much trust it has earned over time. It is designed to be:
  • workspace-friendly (enterprise backend teams can make fast decisions)
  • explainable (facets + signals, not a single opaque score)
  • hard to game (signed feedback, time decay, anomaly detection)

​
Why reputation exists

Agent identity + signatures answer: “who is calling?” Reputation adds:
  • “how has this agent behaved historically?”
  • “should I rate-limit, require extra verification, or block?”
This is especially useful for:
  • paid APIs
  • financial actions
  • high-impact automation (deploys, deletes, transfers)

​
Facets (v1)

Rather than a single global score, Clawb returns a small set of facets.

​
verification_tier

How strongly the agent’s identity is established. Example tiers:
  • unverified (new, not attested)
  • verified (attested key ownership)
  • partner_verified (additional checks / enterprise attestation)

​
risk_tier

A coarse risk bucket for decisioning. Example:
  • low, medium, high

​
reputation_band

A maturity signal. Example:
  • new, established, high

​
signals_summary

A compact summary of the inputs (counts, recent windows), suitable for dashboards and workspace decisioning.

​
Signals (inputs)

​
System signals

  • successful vs failed requests
  • policy violations
  • auth failures
  • rate-limit hits
  • anomaly flags

​
Workspace feedback (signed)

Workspace teams using the Verification Service can submit signed feedback signals, e.g.:
  • suspected abuse
  • spam / scraping
  • high error rate
These signals are rate-limited and attributed to the workspace API key.

​
Human/org attestations

Organizations can attach endorsements / attestations for internal agents.

​
Cold start behavior

New agents are treated explicitly as new. Recommended behavior for workspace teams:
  • allow low-risk calls but apply stricter rate limits
  • require higher thresholds for expensive actions
  • step up verification (challenge flows) for sensitive operations

​
Anti-gaming design

  • Workspace feedback is weighted higher than self-reported metrics.
  • Signals decay over time.
  • Anomaly detection flags suspicious bursts.
  • Reputation computations retain provenance (who said what, when).

​
Where you see reputation

  • In the dashboard (agent profile)
  • In verification responses (workspace-facing)
  • In policy evaluation (internal control-plane decisions)
Related: