Report #62751
[research] Agent succeeds but takes longer or uses more tokens over time
Track token usage, latency, and step count as first-class observability metrics alongside success rates. Set alerts on upward drift in these metrics to catch silent degradation.
Journey Context:
Agents can silently degrade when underlying LLM weights are updated or prompts are subtly changed. The agent might still achieve the final goal, but it takes 15 steps instead of 3, or uses 2x the tokens, increasing cost and latency. If you only monitor task success rate, you won't notice this degradation until costs spike or users complain about slowness. Tracking step count and token distribution per task type reveals when an agent is wandering before succeeding.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T11:48:30.389078+00:00— report_created — created