AgentOps Center: running multi-agent systems with OpenTelemetry and SigNoz
A new GitHub project, AgentOps Center, addresses the unique observability challenges of multi-agent AI systems, which can slow down without errors. The project, detailed in a Dev Community article, provides a pattern for tracing agent workflows using OpenTelemetry, custom attributes, and a 'Root Cause Copilot' that gathers telemetry before diagnosis.
Traditional Application Performance Monitoring (APM) often falls short for multi-agent AI systems because these systems can exhibit performance issues, such as endless loops or rate limits, without generating explicit errors. This leads to increased operational costs and delays that go undetected by standard monitoring tools. AgentOps Center proposes a solution by focusing on granular tracing.
The project designs specific telemetry 'spans' for each agent, model call, and tool interaction within a workflow, linking them into a comprehensive trace. This allows operators to pinpoint the exact component causing slowdowns, such as a looping node or a rate-limited tool. It reuses standard OpenTelemetry `gen_ai.*` keys for model calls while introducing custom keys for agent-specific logic, ensuring both broad compatibility and detailed internal visibility.
A notable feature is the 'Root Cause Copilot,' which enhances diagnostic accuracy. Instead of making a 'cold' query to a language model, it first fetches relevant traces, logs, metrics, and alerts from a SigNoz server. This evidence-based approach provides a rich context for the model's analysis, moving beyond simple error detection to address complex, non-error-generating performance bottlenecks inherent in multi-agent architectures.
While the article's author reviewed the code without full deployment, the project demonstrates a practical pattern for monitoring. It shows how to manage high volumes of telemetry data through memory limiting and batching, with reported runs generating thousands of spans. The design also allows for flexibility, supporting self-hosted SigNoz but adaptable to other OpenTelemetry-compatible backends by simply changing the exporter.
Share this article
Related reading
6 stories
AI coding audit trails: what did the agent do in your repo?
We explored the audit trails of AI agents, a key concern when running multi-agent systems.

Probe opened over Nvidia influence on US government

UOB and Guangzhou Municipal Government sign MoU to establish new technology innovation centre in Guangzhou

Baidu fined US$75,000 as China wraps up Joyy antitrust case

Why India needs a trade secrets law and why it needs it now

