Skip to main content
This feature is in early access and behind a feature flag. Contact your Augment Solutions Architect to enable it for your tenant
Trace Export sends a record of what your Cosmos agents do to your own observability platform, using OpenTelemetry (OTel). Use it to:
  • Debug a failed or slow session: see every setup step, model request and tool call in order, with timing and errors.
  • Track usage and cost by model, expert and user.
  • Alert when something breaks, for example when an MCP server’s credentials expire or a scheduled run fails.
OpenTelemetry is a vendor-neutral standard that most observability platforms support, including Datadog, Grafana Cloud, Honeycomb, Langfuse and 20 others. Cosmos follows the OpenTelemetry GenAI and MCP conventions, so these platforms recognize the data without extra setup.

How it works

An organization admin adds a destination: your platform’s OTLP endpoint and API key. From then on, Cosmos sends every session to it as the session runs. Augment doesn’t store the exported data.

Key terms

Which signal do I need?

Each destination always receives traces. Metrics and logs are optional and turned on per destination. Metrics and logs are derived from the same sessions as traces. Turning them on doesn’t record anything extra in Cosmos; it only sends the same activity in a form that’s easier to chart or search.

Get started

1

Open Trace Export

As an organization admin, open Settings → Trace Export.
2

Add a destination

Click Add Destination, pick your platform, and paste its API key. Turn on Send metrics and Send logs if you want them.
3

Test and save

Click Add destination. Cosmos sends a test span and saves and enables the destination when it’s accepted.
4

Run a session

Start any session, then search your platform for service.name = augment-cosmos.
For every option, see Set up Trace Export.

What a session looks like

This example follows one session in which a user asks an expert to fix a flaky test, in a fresh cloud environment with a Linear MCP server.
  1. The session starts. Cosmos starts the environment (runtime_acquisition), runs its startup script (startup_script) and connects the Linear MCP server (initialize linear).
  2. The user’s message starts a run (invoke_agent Flaky Test Fixer, named after the expert). Its wait_for_agent span covers the time until the agent received the message.
  3. The agent’s turn (invoke_agent auggie) asks the model what to do (chat Claude Sonnet 4.5, named as in the Cosmos model picker).
  4. The agent reads a file (execute_tool view) and runs the tests (execute_tool launch-process), then asks the model again.
  5. The agent delegates a search to a subagent. The tool call that starts it (execute_tool agent) is the parent of the subagent’s span (invoke_agent explore), and the subagent’s own tool calls sit under that. Subagents’ model requests aren’t traced yet, so their usage and cost aren’t in the trace.
  6. The agent replies and the run ends. When the session is archived, the session span ends.
Durations are illustrative. Things to notice:
  • Setup time is visible: the user waited 1 minute for wait_for_agent before the agent started working.
  • Every span has gen_ai.conversation.id, the session ID, so you can find the whole session from any one span.
  • Model requests carry usage and cost: token counts (gen_ai.usage.*), Augment credits (cosmos.credits), and cost in US dollars for dollar-billed organizations (cosmos.cost_usd). gen_ai.response.model is the public model ID, such as claude-sonnet-4-5, so platforms that price LLM calls can price these.
  • A failure has a reason: a failed span has status ERROR and an error.type, such as mcp_auth on initialize linear if the Linear credentials had expired. Setup steps carry the expert and the trigger, so you can tell it was this expert’s scheduled run that broke.
For how to read and query these spans, see Use your data. For every span and attribute, see the Export reference.

Content and privacy

By default, Cosmos sends only metadata. An admin can turn on three content levels: Conversation, Model inputs and Tool inputs and outputs. Before anything leaves Cosmos, recognized credentials are redacted and private model reasoning is removed. See Content and privacy.

Who can use it

  • Plans: Enterprise and team organizations. Not available on individual plans.
  • Role: organization admins configure it. Other members don’t see the Trace Export page.

Availability

Trace Export is in early access and behind a feature flag. Contact your Augment Solutions Architect to enable it for your tenant. The page, the supported providers and the exported data can change before general availability. Not supported yet:
  • OTLP/gRPC and OTLP/JSON. Cosmos sends OTLP/HTTP with protobuf only.
  • Configurable sampling. There’s no sampling setting. Cosmos sends every span it captures, but the agent stops creating new spans for a run after 2,000, so very long runs can be missing later model requests and tool calls (and their usage). To reduce volume, export only selected experts, or sample in your own OpenTelemetry Collector.
  • Keyless cloud authentication. Destinations authenticate with request headers. Google Cloud Trace, AWS SigV4 and Azure Monitor identity federation aren’t available.
  • Non-OTLP destinations, such as Amazon EventBridge, webhooks or object storage. Send to an OpenTelemetry Collector and forward from there.
  • Replay. Augment doesn’t store exported data, so data that was dropped, or produced while a destination was disabled, can’t be sent again.
  • Joining your application’s traces. Each Cosmos session starts its own trace; you can’t pass in a trace context to make a session part of a trace in your own application. In the other direction, your MCP servers can join Cosmos traces.
  • A truncation marker. When content is shortened to fit, the span carries no attribute saying so.
  • An API or CLI for destinations. Destinations are managed on the Trace Export page only.

Next steps

Set up Trace Export

Add a destination and choose what to send.

Use your data

Find sessions, count tokens and cost, build dashboards and alerts.

Content and privacy

What’s sent, what’s never sent, and how content is limited.

Export reference

Every span, attribute, metric, log event and error type.