ANTHROPIC SHIPS OTEL TRACE EXPORT — TRACEABILITY IS THE NEW BASELINE

Is your AI agent actually traceable?

Anthropic just added OpenTelemetry trace export to the Claude Platform — agent traces are becoming the industry standard. Meanwhile the EU AI Office is demanding evidence of agent behavior. If your agent went rogue tonight, could you answer three questions: who called what, which tools fired, and where did the data go? Run this free 5-point self-check. No signup, nothing installed.

The 5-point self-check

Can you name every tool your agent called this week? Pick any agent and list every tool invocation — function name, arguments, timestamp. Not a guess: the record.

Fail: you can name the tools the agent has access to, not which ones it actually called.

Pass: a complete per-agent tool-call record, queryable by date.

Can you trace one data flow end to end? Take a sensitive input — a customer record, a secret — and trace every step where your agent touched it: retrieved, transformed, sent where?

Fail: the trail goes cold after the first tool call.

Pass: full lineage: input → transformations → destinations.

Do your traces follow a standard? OpenTelemetry is now the lingua franca — Anthropic exports OTel spans. Ad-hoc JSON blobs in a vendor-locked format aren't interoperable evidence.

Fail: traces in a proprietary format nobody else can read.

Pass: OTel-compatible spans, exportable and portable.

Can you reconstruct a session after the fact? Take an agent session from two weeks ago. Can you replay exactly what happened, in order, with outcomes?

Fail: retention expired, or sessions are unreconstructable.

Pass: ordered, replayable session history on demand.

Would the trace convince a skeptic? If a regulator, a customer, or your own security team challenged one trace, could you verify its integrity — or is it your word against theirs?

Fail: unverifiable traces; trust me, not proof.

Pass: tamper-evident traces a third party can verify.

If you failed any check, do this now

Instrument every agent — record every tool call, argument, and outcome from day one.
Emit standard traces — OTel-compatible spans, so your evidence works everywhere.
Keep session history — replayable records, retained long enough to matter.
Make traces verifiable — hash-chain them so integrity is provable, not promised.

Make your agents traceable → Agent Flight Recorder

This page covered the 5 checks. Agent Flight Recorder keeps a hash-chained, tamper-evident log of every agent action with OTel-compatible traces.