Project · 1 min read

Closed Loop

July 1, 2026

LLM agent workloads fail in ways request-level APM doesn't surface — runaway token spend, tool-call loops, and failure modes that live between spans rather than inside them.

Why this exists

Built at the Agents of SigNoz Hackathon (2026), as a team of three. The question we wanted to answer: how much agent-specific observability can you get from stock SigNoz — no forks, no custom UI — if you shape the telemetry right on the way in?

What it does

• Agent-level tracing — traces structured around agent runs and tool calls, not just HTTP requests, so a reasoning loop reads as one story.
• Cost views — token spend surfaced per agent and per run, because cost is a first-class failure mode for agentic systems.
• Failure-mode views — the ways agents actually break (loops, tool errors, dead-end runs) made visible instead of buried in span attributes.

How it runs

Stock SigNoz over a ClickHouse-backed OTLP pipeline, self-hosted on Azure behind nginx with TLS termination.