OpenAI Agents SDK observability
Trace each agent decision in production
Instrument OpenAI Agents with Anosys and see spans, tool calls, cost, and quality context from your first production traces.
Questions this answers
- Debug bad tool calls and failed handoffs
- Find slow spans and stalled workflows
- Detect regressions after prompt or model changes
- Attribute spend to agents and outcomes
- Route incidents with trace-level evidence
Follow each agent run from prompt to handoff: spans, model calls, tools, retries, evals, release context, users, cost, and operational outcomes.
Agent runs, turns, handoffs, and spans
Prompts, completions, model parameters, and tool calls
Latency, retries, errors, token usage, and cost
Eval scores, policy checks, and release context
User, workflow, customer, and business KPI context
Start with the docs
Use the linked guide to install the collector or SDK, verify the endpoint, and send a small trace or event payload. Once data lands, connect it to dashboards, evals, alerts, and investigations.
from anosys import trace
with trace("checkout-agent"):
result = await agent.run(input)Where to go next
Does this require OpenTelemetry?
No. You can instrument through the Anosys SDK, OpenTelemetry, or REST depending on your stack.
Can I redact content?
Yes. Sensitive payloads can be controlled with redaction and enterprise data handling policies.
Can I track cost?
Yes. Anosys links token usage and model cost to agents, tools, users, workflows, and outcomes.
Can I monitor tool calls?
Yes. Tool calls are first-class spans with latency, success, failure, and result context.
Can I use this with enterprise security controls?
Yes. Enterprise plans support governance rules, audit trails, ownership, and sensitive data encryption controls.