Agent traces
agentbreeder traces lists each deployment's recent agent traces from the stack's Phoenix: models, tokens, tool calls and latency, never message content.
Preview
Run on macOS with Docker Desktop against Northwind and a live Phoenix (2026-11-15): traces, a
single trace's tree, the console's Traces tab, Try it's trace spans and an eval case's trace.
After a fix, trace token totals equalled the runtime's count and LiteLLM's spend log.
agentbreeder traces lists each deployment's recent agent traces from the stack's Phoenix: models, tokens, tool calls and latency, never message content.
agentbreeder traces --stack . # every deployment not removed or failed
agentbreeder traces ID --stack . --last 50 # one deployment
agentbreeder traces --trace TRACE_ID --stack . # one trace, as a treetraces reads the stack's service/phoenix on its loopback port. Agents built from agent.yaml
with traces: otlp send OpenInference spans there. From runtime/adk@1.2.0, a trace is tied to
its deployment by the deployment ID on its root span; older runtimes are matched by time window.
The ID is reported by the agent, not authenticated.
Shown: span names and kinds, status, latency, models, token counts (each model call counted
once), tool names, error messages, and from runtime/adk@1.3.0 a gateway fallback
(fallback: ALIAS).
Never shown: prompts, replies, tool arguments and results, and retrieved documents, because they hold customer messages. Every string is redacted with the stack's stored secrets; if a secret can't be read, nothing is printed.
service/phoenix@1.1.0 mints a read-only key for this. On phoenix@1.0.x, traces falls back
to the ingest key, which is admin-equivalent in Phoenix, and says so.
In the console
An agent's Traces tab lists one deployment's traces, and each span tree opens on request. Try it opens the trace of a turn, and the eval report viewer opens a case's trace. In Try it, a trace read before Phoenix has all its spans is marked as still arriving, with Read again.
What traces don't cover yet
- Message content (
--content). - Eval-run traces linked from
releasein the CLI.
Canonical reference: traces.md.
LLM cost
agentbreeder cost reports LLM spend per deployment and per model from the gateway's LiteLLM spend logs. An estimate from list prices, not a bill.
Identity gate
agentbreeder deploy refuses an agent that is not onboarded in your identity provider. Policy, approvals, a kill switch, the audit log and the MCP egress sidecar.