Export agent telemetry

Inspect A2A operations in your OpenTelemetry backend and connect them to model and tool activity.

Agent Logs is the dashboard view of gateway activity. Observability connectors send completed traces to your monitoring backend. Use the OpenTelemetry connector when you want to inspect agent traffic alongside the rest of your service’s traces.

Configure OpenTelemetry

Follow the OpenTelemetry guide to connect your collector. For deployment configuration, see the config.json plugin settings or Helm plugin settings.

Find an agent operation

A2A operation spans carry attributes that identify the work:

Attribute	What it identifies
gen_ai.agent.name	Registered agent name.
bifrost.a2a.operation.name	Gateway operation, including protocol and push operations.
a2a.method.name	A2A v1 JSON-RPC method, when applicable.
a2a.task.id	Task ID supplied in the request, when available.
gen_ai.conversation.id	Context ID supplied in a message request, when available.
bifrost.a2a.upstream.transport	Transport Bifrost used to call the agent.
error.type	Error classification for a failed operation.

Look for the operation and IDs relevant to your request. The first message often creates its task and context IDs in the response, so that operation’s span can lack both IDs. Subsequent message requests can carry both; task operations such as GetTask carry the requested task ID. Discovery and early failures may also lack these IDs.

For example, a successful JSON-RPC message produces an operation span like this. The values below show the fields to look for, with a sample agent name:

span: a2a.send_message.library-research
kind: CLIENT
status: OK
gen_ai.operation.name: invoke_agent
gen_ai.agent.name: library-research
bifrost.a2a.operation.name: SendMessage
a2a.method.name: SendMessage
bifrost.a2a.upstream.transport: JSONRPC

The connector also records a2a.client.operation.duration in seconds, with operation and available transport, method, error, and governance dimensions. This measures the gateway operation; it is not the model’s token-generation metric.

Connect model and tool activity

Pass the context and task correlation headers when your agent calls Bifrost for LLM or MCP work. Session grouping can place requests sharing a session ID in one trace when no explicit incoming trace context takes precedence.

For a full distributed trace across your application, agent server, and Bifrost, instrument those services and propagate W3C trace context. Matching session IDs help group a conversation, but do not by themselves describe parent-child relationships across every service.

Why matching IDs do not always produce one trace

The A2A context ID and x-bf-session-id have separate jobs. The context ID identifies the A2A conversation. The session header tells the connector which requests to group. To group A2A calls with your agent’s LLM and MCP calls, send the same session header on those requests and enable group_traces_by_session.

When an incoming traceparent identifies an existing distributed trace, Bifrost keeps that trace instead of assigning a trace from the session ID. Filter by session.id to find the related requests even when they belong to different distributed traces.

Dashboard raw capture and connector payload export are separate settings. Enabling Agent Gateway raw capture does not enable every connector’s content export. Review the connector’s content and attribute controls before exporting conversation data.

Other connectors have different output formats and configuration. See the Datadog, Maxim AI, or Kafka guide for the connector you use and the signals it supports.

Use Troubleshooting to diagnose authentication, transport, callback, or correlation failures.