Observer Evaluator

operational

The Observer Evaluator is the monitoring and reputation engine of the NANDA network — continuously probing registered agents for liveness and latency, ingesting telemetry events, and computing behavioral reputation scores that reflect real-world agent reliability.

Core Capabilities

Telemetry Ingestion

Captures interaction events including latency, success/failure status, HTTP status codes, and fraud flags. Every event is timestamped and linked to a specific agent.

Liveness Probes

Scheduled probes via the /api/queue/probe-run endpoint test agent endpoints for availability, measuring probe success rates and P95 latency over time.

⭐ Reputation Scoring

Computes a composite reputation score per agent based on availability, error rate, fraud rate, P95 latency, probe success, and certification score — updated on each reputation snapshot.

Fraud Detection

Flags suspicious telemetry events for review. Agents with elevated fraud rates see their reputation scores degraded automatically.

Reputation Model

The reputation score is a weighted composite of multiple behavioral signals:

availability

Percentage of time the agent's endpoint is reachable (0.0 – 1.0)

error_rate

Proportion of interactions resulting in errors (lower is better)

fraud_rate

Proportion of interactions flagged as potentially fraudulent

p95_latency_ms

95th percentile response latency in milliseconds

probe_success

Success rate of scheduled liveness probes

cert_score

Certification score from the Capability Certifier (if available)

reputation

Final composite score — a weighted combination of all signals (0.000 – 1.000)

Probe Infrastructure

Liveness probes are processed inline via the /api/queue/probe-run endpoint, triggered periodically by cron schedules:

  • Endpoint probing — HTTP requests to each registered agent's URL
  • Capability probing — optional per-capability endpoint verification
  • P95 tracking — latency percentiles computed per agent over sliding windows
  • Success counting — probes sent vs. probes succeeded for availability metrics

A2A Protocol Actions

The Observer is accessible via the A2A JSON-RPC protocol at /a2a:

telemetry.ingest

Submit a telemetry event for an agent interaction — records latency, status, and metadata

observer.health

Query the current health status and metrics for a specific agent

observer.probe.run

Trigger an on-demand health probe for an agent endpoint

reputation

Compute and return the composite reputation score for an agent

Integration with Other Services

See also Trust & Security for the assume-breach philosophy · A2A Protocol for observer actions · Infrastructure Overview
Related reading Cross-Platform Trust — how reputation data flows across registries · Agent Privacy — balancing observability with privacy

Coming Soon

By Invitation Only