Observability data takes many forms: logs, metrics, traces, and now spans from AI agents. Each has traditionally meant a separate collector, a separate store, and a separate query language. The complexity is structural, and so is the cost.

OpenTelemetry is a CNCF-backed, vendor-neutral standard for generating, collecting, and exporting traces, metrics, and logs. Instrument once and retain the flexibility to switch or evolve your backend without re-instrumenting your code.

OpenSearch ingests OTLP data via the OTel Collector on standard ports, 4317 for gRPC and 4318 for HTTP. Instrumentation uses standard OTel SDKs. There are no proprietary OpenSearch libraries required.

From the Collector, the pipeline forks by signal: metrics route to Prometheus for time-series storage, while traces and logs route to OpenSearch Data Prepper, which enriches them before writing them to OpenSearch. That enrichment step builds the service map, derives RED metrics from span data and writes them to Prometheus, and preserves the trace IDs that link logs to the requests that produced them. In Dashboards, a trace’s related logs are one tab away.

The Observability Stack ships this path preconfigured, by installer, Docker Compose, or Helm.

Depicts OpenSearch Dashboards 3.8

What the pipeline produces is covered in greater depth on the following pages: service maps and trace waterfalls in Application Performance Monitoring, PPL and log correlation in Log Analytics, PromQL and dashboards in Metrics and Prometheus.

Depicts OpenSearch Dashboards 3.8

Instead of relying on proprietary schemas like most agent observability tools, OpenSearch has adopted the OpenTelemetry GenAI semantic conventions. This allows agent invocations, tool calls, and LLM interactions to be described with attributes anyone can read and any OTel-compatible backend can consume.

If your agents are already instrumented with something else, the stack converts on the way in. The Collector configuration shipped with the Observability Stack enables the upstream OpenTelemetry GenAI normalizer, which maps OpenInference and OpenLLMetry attributes into GenAI semantic conventions before AI agent traces are stored.

In production, agent telemetry lands alongside your service traces and is explored in OpenSearch Dashboards, where an invocation renders as a timeline and as a directed graph. Agent Health is the local companion to that, an npx tool for inspecting traces, running benchmarks, and scoring outputs before anything ships.

Agent observability in OpenSearch Dashboards
Timeline and graph views for agent traces. Agent Health adds local benchmarking and evaluation.

Instrumented where agent teams already work
OpenAI · Anthropic · LangChain · LlamaIndex · Cohere · Mistral · Groq · Ollama · Google · Amazon Bedrock · Strands Agents SDK

By combining Apache 2.0 projects, OpenTelemetry and OpenSearch remove licensing barriers for everything on your observability pathway. The Collector configuration, the ingestion layer, the storage layer, and the query and visualization layer are all open. No proprietary tiers, no features gated behind a license, and no vendor lock-in.