trace.init(), Veritrix silently instruments your agent runtime without any additional code. Every LLM call, tool execution, and agent handoff produces a structured span that appears in your Veritrix dashboard in real time. There is nothing else to configure.
Auto-instrumented span types
Veritrix captures four span types out of the box. You do not need to create, open, or close spans manually.llm.call — LLM invocations
llm.call — LLM invocations
Emitted for every call made to a large language model. Captures the full
prompt, the completion, input and output token counts, wall-clock latency,
and the final status of the call.
tool.exec — Tool and function calls
tool.exec — Tool and function calls
Emitted whenever an agent invokes a tool or function (web search, code
execution, database lookup, etc.). Captures the tool name, the input
arguments, the return value, and latency.
agent.handoff — Agent-to-agent transfers
agent.handoff — Agent-to-agent transfers
Emitted when one agent passes control to another — for example, a
coordinator delegating a subtask to a specialist agent. Captures the source
agent, the target agent, and the payload passed between them.
llm.retry — Retried LLM calls
llm.retry — Retried LLM calls
Emitted for each retry of a failed LLM call. Captures the retry count,
the error that triggered the retry, and the final outcome so you can
diagnose reliability issues without noise in your
llm.call spans.No manual span creation is needed. Veritrix instruments your agent
automatically as long as
trace.init() has been called before any agent
code runs.Span payload fields
Every span — regardless of type — carries the following fields.Example span payload
The JSON below is a realllm.call span from a reviewer agent that failed to parse a model response.