OpenAI Agent Telemetry

Measurement path

Connect detailed execution to a durable product outcome

Keep step-level traces in the specialist tracing path, then emit one compact terminal event that can join agent quality, cost, reliability, and product behavior.

1
Agent run
Start with an approved run identifier, workflow, model, and prompt version.
2
Tools and retries
Retain detailed steps in tracing and categorize failures without copying payloads.
3
Terminal outcome
Emit success, failure, cancellation, or human handoff with cost and duration.
4
SQL decision
Compare accepted outcomes, failures, and unit economics by feature and release.

Before you start

Prerequisites and boundaries

The OpenAI Agents SDK and telemetry-sh packages initialized in a server runtime
A documented workflow name and terminal success, failure, and handoff outcomes
A decision about which prompt, tool-input, and output content must remain excluded

Delivery setup

Install and initialize server-side

Import telemetry-sh in server-only code and initialize it once with process.env.TELEMETRY_API_KEY. Keep ingestion credentials out of browser bundles, client-visible environment variables, source control, logs, and exception messages.

openai-agents-install

npm installation

bash

npm install telemetry-sh

1Prepare one reusable server-side delivery client with bounded network behavior.
2Add the outcome event at the success, failure, retry, or timeout boundary.
3Send controlled fixtures and inspect the stored rows before enabling an alert.

Snippet

Start with one structured event

Add this shape where the workflow completes, fails, or retries. Then build the dashboard from real fields.

openai-agents

OpenAI Agent Telemetry event

javascript

await telemetry.log("agent_tool_called", {
  run_id: runId,
  workflow: "support_resolution",
  agent_name: "support_agent",
  model: "gpt-4.1",
  tool_name: "lookup_order",
  status: "success",
  latency_ms: 842,
  retry_count: 0,
  estimated_cost_usd: 0.018,
  prompt_version: "support-v3",
  release: process.env.APP_RELEASE,
});

Event contract

run_id, workflow, agent_name, model, and prompt_version

status, duration_ms, retry_count, total_tokens, and estimated_cost_usd

tool_name, tool_status, human_handoff, reviewer_outcome, and release

Implementation checkpoints

Checkpoint 1

Emit one compact product outcome when the complete run finishes; keep SDK traces for step-level debugging instead of copying every span into a second system.

Checkpoint 2

Use a shared run_id or approved trace_id to connect separate tool events without storing tool arguments or results.

Checkpoint 3

Exercise tool failure, guardrail rejection, handoff, cancellation, and exporter-flush behavior before relying on completion-rate dashboards.

Verification

Prove the event arrived

Run this after exercising known success and failure cases. Replace the fallback table name if your final event contract differs from the snippet.

openai-agents-verification

OpenAI Agent Telemetry verification query

sql

SELECT *
FROM agent_tool_called
ORDER BY timestamp_utc DESC
LIMIT 20;

Confirm one terminal row per logical outcome, with the expected status, identifiers, units, and UTC time.

Inspect the inferred schema and verify that retries do not change field types or generate a new logical event ID.

Search the stored fields for credentials, raw payloads, prompts, private content, and unbounded error messages.

Exercise a provider timeout, ingestion rejection, and process shutdown before treating the dashboard as complete.

Implementation references

Review the event contract, data-safety guidance, and upstream primary documentation before enabling a new production path.

AI agent quality evaluation Redacting sensitive data Nested tool-call SQL OpenAI Agents SDK tracing

Production boundary

Keep the outcome event small and recoverable

This pattern provides

A bounded, SQL-ready outcome beside the upstream workflow.
Stable fields for dashboards, alerts, and cross-event correlation.
A fixture-driven path for validating success, failure, retry, and timeout behavior.

This pattern does not provide

An OTLP exporter, automatic collection pipeline, or replacement for detailed traces and diagnostic logs.
Exactly-once delivery merely because the payload contains an event ID.
Permission to collect raw provider payloads, user content, credentials, or regulated data.

Delivery and idempotency Batching and shutdown Troubleshoot ingestion

Event schema starting points

Review the row grain, emit boundary, required types, privacy classes, example payload, and validation checklist before adapting a query or snippet to production.

AI operations10 fields