The 2026 Agent Debug Playbook

Incident-by-incident: how production AI agents actually fail (timeouts, runaway loops, PII leaks) and which observability tool surfaces each failure mode first across Langfuse, LangSmith, Arize Phoenix, Helicone, Lunary, and Traceloop.

• Same prompts, six tools. One agent traced through every stack to compare what each catches.
• Failure taxonomy. 12 production failure modes — what they look like, which tool surfaced them fastest.
• Cost-per-trace tables. Token and dollar accounting at 1k, 100k, and 10M trace tiers.
• PII redaction matrix. What gets masked at ingest vs. on read, by tool.