drift-review

Drift Review — PromptSpeak Governance

You are conducting a behavioral drift review across governed agents.

Workflow

1. Get system overview — Call ps_state_system to see all agents, their circuit breaker status, and drift metrics.
2. For each agent with drift above baseline:
   • Call ps_state_drift_history with the agent ID
   • Identify the drift trend (increasing, stable, decreasing)
   • Note which dimensions are drifting (goal displacement, tool usage patterns, confidence shifts)
3. Present findings as a dashboard:

Agent: [agent_id]
  Status: [circuit breaker state]
  Drift: [current drift score] (baseline: [baseline])
  Trend: [↑ increasing / → stable / ↓ decreasing]
  Top concern: [most drifted dimension]

4. Recommend actions based on severity:
   • Low drift (< 0.3 above baseline): Monitor, no action needed
   • Medium drift (0.3–0.7 above baseline): Suggest recalibrating baseline with ps_state_reset
   • High drift (> 0.7 above baseline): Recommend halting agent with ps_state_halt until root cause is identified
5. If the user wants to act, execute the recommended tools.

Guidelines

• Always show data before recommending actions.
• Drift is not inherently bad — agents learning new tasks will drift. Context matters.
• Check audit logs with ps_audit_get if the user wants to understand what caused the drift.
• A halted agent (circuit breaker open) should be investigated before resuming with ps_state_resume.