Recommendations

Recommendations is a rule-based pattern detector, not a model — three fixed SQL-aggregate checks run against the last 24 hours of invocation and decision-log data in a single pass. A background scheduler runs the pass every 24 hours (with an initial run 60 seconds after boot so a freshly deployed instance has findings within a minute), and an "Analyze now" button on the tab triggers the same pass on demand. Each run clears out all non-dismissed recommendations first and re-inserts whatever it currently finds, so the list always reflects the latest pass rather than accumulating stale findings; dismissed recommendations are left alone.

The three detectors are: redundant calls, where an (agent, tool) pair has made more than 50 calls in 24 hours and over 70% of them share identical arguments (via an args_hash on the invocation) — severity scales with how redundant the calls are, and the finding estimates wasted tokens from the gap between total and unique- argument calls; retry loop, where an (agent, tool) pair has racked up more than 10 denied calls in 24 hours, a strong signal the agent's error handling blindly retries on deny instead of stopping; and high session density, where an agent's average calls-per-hour is more than 3x the median across all other agents in the window (with a floor of 20 calls/hour to avoid flagging low-traffic noise) — a possible reasoning loop or inefficient tool-use pattern.

Each recommendation card shows a severity-colored left border (red for high, amber for medium, slate for low), the finding and suggested fix as plain text generated directly by the detecting query, and — for redundant-call findings — an estimated token-waste figure. Dismissing a card records an operator-supplied reason and removes it from the active list until a future pass re-detects the same pattern.