Five production AI agent failure patterns and the telemetry that catches them

Bonree · reddit · 2026-11-12

The author (a Bonree employee, disclosed) catalogs five failure patterns that recur when AI agents move from demo to production, plus the observability needed to catch each:

Silent wrong answers: a tool fails inside a "successful" response and the agent answers confidently anyway. → Record tool outcomes per step, separate from transport status; flag runs where a final answer follows a failed/empty tool call.
Loops and runaway retries: near-identical tool calls repeat, cost drifts upward. → Track step counts, repeated argument signatures, alert on outlier run length.
Latency with no owner: model, retrieval and downstream times each look fine individually. → Split model/tool/dependency time on a single trace.
Cost growing faster than traffic: context accumulates, large tool results passed back in full. → Track input tokens per step and tool result size by agent and tool.
Irreproducible failures: nobody recorded which model/prompt/tool versions were involved. → Attach version metadata to every run.

The common thread: agent observability must work per step, not per request, and share one trace with infrastructure telemetry. The author also describes Sage AI, an AIOps agent in Bonree ONE that assists diagnosis on the same telemetry, and invites others to compare notes.

Original post →

Five production AI agent failure patterns and the telemetry that catches them

More from coding & agent