OpenTelemetry standardizes signals and pipelines; it does not create missing observations
OpenTelemetry provides APIs, SDKs, instrumentation libraries, OTLP, collectors, exporters, resources, and semantic conventions across traces, metrics, logs, and profiles. The Collector can receive, process, and export telemetry to one or more backends. Those pieces make independently produced telemetry interoperable, but a semantic convention cannot describe a child process, file operation, or local network effect unless some measurement source actually observes and emits it.
AgentSight starts from a different measurement boundary
AgentSight v1.0.31 can combine native agent-session parsing with Linux process/file/resource capture and supported TLS/plaintext reconstruction. A saved local session can therefore contain model activity beside process trees, path effects, network targets, resource phases, tool/workflow rows, and provenance/confidence fields. Those records are useful precisely because they do not depend on every agent or descendant being instrumented with an OpenTelemetry SDK.
The v1.0.31 exporter accepts completed materialized LLM calls, not arbitrary session rows
The current OtelExporter is attached to the materialized llm_call view. A call without an end timestamp is skipped; each completed call becomes one CLIENT span named chat <model>. The exporter maps supported gen_ai request/response fields, provider and server address, token usage, finish reasons, and HTTP status. Prompt and completion content remain opt-in through --otel-capture-content. This is a projection after AgentSight has reconstructed and correlated a model exchange, not a direct serialization of raw eBPF events.
Most AgentSight system context does not automatically become GenAI spans
Current v1.0.31 documentation explicitly does not emit execute_tool, invoke_agent, invoke_workflow, or plan spans. Process, file, network, and resource rows are not automatically mapped by this GenAI sink, and AgentSight-specific provenance/confidence stays in AgentSight rows. If a central OTel backend needs those dimensions, treat that as a separate instrumentation or export design rather than assuming debug trace --otel already transports the full session.
Trace grouping is useful but intentionally does not invent a call tree
AgentSight groups exported calls by an explicit conversation or thread identifier when one is available, then by AgentSight session ID, then by a recording-scoped fallback trace. Each call receives a unique span ID, but v1.0.31 does not infer root/child relationships between calls. Preserve the AgentSight service/source identity when native agent spans and AgentSight-derived spans land in the same backend so a shared trace view does not erase how each field was measured.
OTLP delivery is another boundary to verify
In v1.0.31 each eligible LLM call builds an OTLP/HTTP JSON payload and the sink starts an asynchronous POST. Non-success responses and transport errors are logged as warnings, while the current command path does not await every detached export task before process shutdown. A call being present in the local materialized session therefore proves local reconstruction, not collector receipt. When delivery matters, verify the Collector or backend rather than inferring success from the local AgentSight view.
Content export and local capture have different privacy boundaries
AgentSight keeps gen_ai input/output message content off by default for OTel export, even though the local session may already contain sensitive prompts, responses, headers, paths, or network targets depending on capture mode. Sending selected spans to a Collector creates a new storage and access boundary. Apply the Collector and backend's own retention, redaction, and access controls, and do not enable content export merely because local capture exists.
Choose the smallest architecture that answers the question
Use OpenTelemetry alone when instrumented application and service telemetry already answers the debugging or operations question. Add AgentSight when you need local agent-session context, closed-component visibility, descendant execution, or host effects that are outside that instrumentation boundary. Use AgentSight's OTel export when selected model-call spans should join an existing Collector pipeline, while keeping the full local AgentSight session available for deeper system analysis.