LinkedIn publication text
How do you know what an AI agent really did?
Did it finish the task? Did it fail or drift off course? How long did it take and what did it cost? The answer is tracing: every step of the agent becomes a span inside a trace.
I compared 13 platforms (Anthropic, OpenAI, Google, Microsoft, AWS and 7 independents). Versions were read from PyPI and npm on Oct 10, 2026, and features checked against the official documentation. What I found:
🟦 Span, trace and session: a span is one traced step (model call, tool use, handoff); a trace is the whole journey; a session groups several traces from the same conversation.
🟩 A trace is not proof of success: a run with no error only shows the agent did not crash. Content failures often trigger no error status at all, so you need a separate evaluation.
🟧 OpenTelemetry is the path to portability, but the GenAI conventions are still in Development status: names can change.
🟪 Privacy differs by platform: the Claude Agent SDK does not record content by default; the OpenAI Agents SDK records inputs and outputs unless you turn it off.
🟥 Silent failure: on at least one platform, a telemetry export error never shows up. A missing trace may be an export failure.
🟨 Maturity: Claude Agent SDK and Managed Agents traces are in beta; Foundry workflow tracing is in preview; core SDKs are still 0.x.
📌 Bottom line: we are midway through maturation. The solutions are, in large part, preliminary versions, and many are in beta or preview. Span names, attributes and behaviors can change: adopt with the migration cost in mind.
✅ Best practices: instrument with OTel, define in writing what gets captured, monitor the export itself, and keep "finished without error" apart from "correct result".
The full study includes a glossary, 47 references with validation marks and a platform comparison matrix.
Which tracing tool do you use for your agents?