New Relic Adds Tracing for Claude Agents Through OpenTelemetry

What the integration covers

The Claude Platform now exports OpenTelemetry (OTel) traces natively, and New Relic AI Monitoring reads those spans where they land. New Relic says its Normalize on Read architecture queries them straight from the Span table, so nothing is duplicated at ingest. When an application passes its trace context, which the Claude SDK does automatically, each Claude span sits in the same trace as the customer's APM data.

The traces cover three areas, according to New Relic:

• Claude Managed Agents: every session request, agent turn, model request, tool use and outcome evaluation.
• Messages API: each message request with its model requests and compactions, plus any server or MCP connector tool use.
• Claude SDKs: the application's requests and custom tools, for Python, TypeScript, Go, Java, C#, PHP and Ruby.

The spans follow the OTel GenAI semantic conventions, which AI Observability already reads, so there is no attribute mapping to maintain.

Debugging an agent across the whole stack

New Relic describes the problem this addresses: until now, an agent run was split between consoles. The customer's services ran the custom tools, while the Claude Platform ran the model and, with Claude Managed Agents, the agent loop.

The company gives an example of a support agent that takes 40 seconds to issue a refund. AI Observability shows the model replying quickly, then the service spending 36 seconds running the lookup_order tool. That tool called an orders API, which was waiting on a database query that slowed down after a deploy the night before. Because Claude's spans share a trace with APM data, New Relic says, teams can follow the trail past the agent, which it says LLM-only tools cannot do.

The product offers several views for this work:

• Agent Trace Waterfall: model requests and tool uses on one timeline inside the customer's own service spans.
• Agents as Entities: each Claude Managed Agents agent gets its own entity, with latency and error rate.
• Conversation Replay: step through a session message by message, with prompts and responses beside timing when Full content is on.
• Alerts and Dashboards: NRQL alerts and existing dashboards work on the data.

Governing tool use

With Full content enabled, each tool span carries the tool's arguments and result, giving teams a record of what each agent touched. Customers can instead choose Metadata only, which keeps timing and identifiers, names.

The same spans can feed New Relic AI Evaluation, which is in public preview. With Full content turned on, it screens live Claude prompts and completions for prompt injection, PII leakage and toxicity.

Setup and availability

An organization admin adds New Relic as a destination on the OpenTelemetry settings page of the Claude Console, picks a New Relic region (US, EU or JP), enters an ingest license key, and chooses which workspaces send traces. Teams also set up tracing in the Claude SDK, as Anthropic's OpenTelemetry guide shows, and point the application's OpenTelemetry exporter at New Relic. According to New Relic, spans arrive within about a minute, with no code changes on the API side.

Claude Platform tracing is in public beta for the first-party Claude Platform. New Relic says support for Amazon Bedrock and Google Cloud Vertex AI is coming soon.

What it means for teams running Claude agents

For teams that run Claude agents alongside their own services, the change brings model behavior and application performance into one trace instead of separate consoles. Slow tool calls, downstream APIs and databases can be examined in the same view as the model requests. Teams that need a record of agent activity or content screening can choose between Full content and Metadata only, depending on how much prompt and response data they want to capture.