Use case

Catch silent failures in agent runs

Multi-step agents fail quietly — a rate limit here, a tool error there, a retry loop burning tokens. Puse traces every step and alerts when error rates climb.

Instrument your agent

The problem

Agent workflows chain dozens of model calls. A provider hiccup mid-loop doesn't always surface to users immediately — but retry logic keeps firing, costs climb, and latency spikes. Existing observability tools show you traces; Puse pages you when reliability regresses.

How Puse helps

  • Every model call traced — including streaming tool-use responses
  • Error messages preserved with model and latency context
  • Slack alerts on error-rate spikes over configurable windows
  • Per-project health charts to spot regressions before deploy review

Example alert

🔴 Error rate spike — project "support-agent"
   12.4% errors (baseline 1.2%) over last 15m
   Top error: rate_limit_exceeded (anthropic)
   p95 latency: 4.2s (↑ 180%)
← All use cases