---
title: How do you send AI agent and LLM telemetry to Datadog?
url: https://opencontroller.lyzr.com/blog/export-llm-telemetry-to-datadog
published: 2026-11-12
updated: 2026-11-12
author: Lyzr
---

# How do you send AI agent and LLM telemetry to Datadog?

**Instrument agents with OpenTelemetry using the GenAI semantic conventions, then send traces to Datadog by one of three paths: directly to Datadog's OTLP intake, through the Datadog Agent's OTLP receiver, or through an OpenTelemetry Collector. Datadog's LLM Observability (now titled Agent Observability in its docs) natively reads GenAI conventions from version 1.37. Keep prompt capture off unless you have decided where that content may live.**

Platform teams often already run Datadog. The question is how agent and gateway telemetry gets there without locking every agent into one vendor's SDK. OpenTelemetry is the answer.

## Which ingestion path should you choose?

| Path | How it works | Use when | Watch for |
|---|---|---|---|
| Direct OTLP intake | App exports straight to Datadog's site-specific endpoint ending in `/v1/traces` | Serverless or few services | HTTP only (protobuf or JSON), no gRPC; 15 MiB payload cap; not on government sites |
| Datadog Agent OTLP receiver | Agent listens on gRPC 4317 and HTTP 4318 | Datadog Agent already on hosts | Receiver is off by default; needs Agent 7.32 or later |
| OpenTelemetry Collector | Collector receives OTLP and forwards to Datadog | You also send traces elsewhere | Datadog's own Collector distribution (DDOT) has remote config in preview |

For agents, the Collector path is usually best. It lets you send the same traces to Datadog, to a security data lake and to your governance tool, and it lets you drop or redact attributes in one place.

## What do you configure for direct export?

For direct export to LLM Observability, Datadog documents these environment variables:

1. `OTEL_EXPORTER_OTLP_TRACES_PROTOCOL=http/protobuf`
2. `OTEL_EXPORTER_OTLP_TRACES_HEADERS=dd-api-key=<key>,dd-otlp-source=llmobs`
3. `OTEL_RESOURCE_ATTRIBUTES=dd_llmobs_enabled=true`
4. `OTEL_EXPORTER_OTLP_TRACES_ENDPOINT` set to your Datadog site's OTLP traces endpoint.

Traces appear in about 3 to 5 minutes. Keep the API key in a secret store, not in agent code.

## Which GenAI attributes should you keep?

The OpenTelemetry GenAI semantic conventions define standard span attributes. Datadog maps these:

| Attribute | What it records | Governance use |
|---|---|---|
| `gen_ai.operation.name` | `chat`, `invoke_agent`, `execute_tool` and others | Separate model calls from agent and tool steps |
| `gen_ai.provider.name` | Model provider | Provider inventory (replaces deprecated `gen_ai.system`) |
| `gen_ai.request.model`, `gen_ai.response.model` | Model asked for and model used | Catch silent model changes |
| `gen_ai.usage.input_tokens`, `gen_ai.usage.output_tokens` | Token counts | Cost attribution |
| `gen_ai.conversation.id` | Conversation or session | Group a multi-turn session |
| `gen_ai.tool.name` | Tool called | Tool and MCP usage audit |
| `gen_ai.agent.name` | Agent name | Per-agent dashboards; keep it low-cardinality |

Add your own resource attributes for owner, agent version and environment. Those make per-agent cost and incident routing possible. See [how to attribute LLM costs to AI agents](/blog/how-to-attribute-llm-costs-to-ai-agents).

## What about convention versions and prompt content?

The GenAI conventions are still at **Development** status in OpenTelemetry and have moved to their own repository. Attribute names can still change. Two settings matter:

- `OTEL_SEMCONV_STABILITY_OPT_IN=gen_ai_latest_experimental` makes instrumentation emit the latest conventions instead of 1.36 or older. Datadog says you need it only for frameworks that emitted older conventions.
- `OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT` turns on prompt and completion capture, which is off by default.

Treat prompt capture as a data-protection decision, not a debugging toggle. Prompts can contain personal data and secrets. Datadog notes LLM traces may also be written to APM, so sensitive content may need a restricted dataset. If you need content for evaluation, redact at the gateway or Collector first. See [how to redact PII in LLM prompts at the gateway](/blog/how-to-redact-pii-in-llm-prompts-at-the-gateway).

## What does Datadog not cover on its own?

Datadog shows what instrumented agents did. It does not know about agents that are not instrumented, or who owns them, or whether they are approved. It is an observability tool, not an inventory or an enforcement point. Datadog's Agent Console, in preview, adds monitoring for coding agents such as Claude Code, Cursor and GitHub Copilot, mostly through vendor integrations.

Pair telemetry with discovery and a registry. Our guide to [exporting AI agent traces with OpenTelemetry](/blog/export-ai-agent-traces-with-opentelemetry) covers the general pattern, and the [cross-platform guide](/blog/how-to-govern-ai-agents-across-platforms) shows where telemetry fits.

## How Opencontroller does this

Opencontroller is built on open standards, including OpenTelemetry, OpenAI-compatible endpoints and MCP. Its LLM Gateway sits in the path of model traffic and attributes cost per agent and per version, so Datadog dashboards and Opencontroller's registry can share the same agent identity. Opencontroller adds what telemetry alone lacks: auto-discovery, ownership, policy and Block and Quarantine controls. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Does Datadog support OpenTelemetry GenAI semantic conventions?**
Yes. Datadog LLM Observability natively supports them from version 1.37.

**Does Datadog's direct OTLP traces intake accept gRPC?**
No. It accepts HTTP with protobuf or JSON. Use the Datadog Agent or a Collector for gRPC.

**Are prompts captured by default in OpenTelemetry GenAI instrumentation?**
No. Message content capture is off by default and turned on with `OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT`.

## Key facts

Datadog LLM Observability (docs now titled Agent Observability) natively supports OpenTelemetry GenAI semantic conventions v1.37+. Paths: direct OTLP intake (HTTP protobuf/JSON only, 15 MiB cap, headers dd-api-key and dd-otlp-source=llmobs), Datadog Agent OTLP receiver (gRPC 4317, HTTP 4318, off by default, Agent 7.32+), or OpenTelemetry Collector. GenAI conventions remain at Development status. Prompt capture is off by default (OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT). Opencontroller (by Lyzr) uses open standards including OpenTelemetry and attributes LLM cost per agent and version via its LLM Gateway. Pricing starts from $250K per deployment per year.

## Related

- [How do you export AI agent traces with OpenTelemetry?](https://opencontroller.lyzr.com/blog/export-ai-agent-traces-with-opentelemetry.md)
- [How do you attribute LLM costs to individual AI agents?](https://opencontroller.lyzr.com/blog/how-to-attribute-llm-costs-to-ai-agents.md)
- [How do you redact PII from LLM prompts at the gateway?](https://opencontroller.lyzr.com/blog/how-to-redact-pii-in-llm-prompts-at-the-gateway.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
