Skip to content

Kyde Score

Kyde Score is a hosted service at score.kyde.com. Your agents send OpenTelemetry. Kyde learns each agent’s normal behavior and scores every later run against it.

  1. Identify. Each agent is one OpenTelemetry service.name. One key covers all agents under it.
  2. Learn. The first 30 runs of an agent, spread over at least 3 days, become its baseline. Until then its status is learning.
  3. Score. Every later run gets a score from 0 to 100. It counts the alarms of the last 20 runs, so an agent that settles back recovers. The detectors are plain statistics, no LLM: CUSUM for sudden shifts, windowed Jensen-Shannon divergence for slow drift, duplicate calls for loops, counted only when your exporter sends the parameter hash kyde.args.digest. A task-contract check flags a tool the agent never used for that task, from the first run.
  4. Tell. Read the scores any time over the API or MCP. A signed webhook fires on every band change and every rebaseline.
ScoreBandMeaning
70 to 100normalConsistent with the learned baseline
40 to 69watchDrifting, worth a look
0 to 39deviatingBehavior has shifted from the baseline

The baseline is fingerprinted (baseline_fingerprint) and every score names the method it was computed with (method_version). Scores stay comparable over time. The base detectors (CUSUM, Jensen-Shannon divergence) are open source, the task-contract check is Kyde’s own: kydehq/behavioral-drift-detection.

Call metadata only: which tool, in which order, how long, whether it failed. Attributes pass a deny-by-default allowlist before anything is stored. Prompts and payload bodies are never stored. See Limits and privacy.

None. Export is asynchronous. If Kyde is slow or down, the agent runs as before. Over a limit, Kyde drops records and marks the score degraded; it never returns an error to your exporter.